Talking Drupal #558 - Agent Management System

[00:00:00] John: This is Talking Drupal, a weekly chat about web design development from a group of people with one thing in common: we love Drupal. This is episode 558, Agent Management System. On today's show, we're talking about AI, agents, and a system to manage them with our guest, Luke McCormick. We'll also cover AI auto reference as our module of the week Welcome to Talking Drupal.

Our guest today is Luke McCormick. Luke has been building online communities since the earliest days of the web when he founded a startup focused on musicians, photographers, and internet users. Today, he is a San Francisco Bay Area-based Drupal solution architect, organizer of the San Francisco Drupal user group, and one of the organizers of Badcamp, and an AI practitioner exploring how AI agents can be effectively integrated into web development teams.

Luke, welcome to the show, and thanks for joining us.

[00:01:13] Luke: I'm excited to be here.

[00:01:15] John: I'm interested to, uh, just dive into the, um, uh, uh, startup. Uh, sorry, the, the musicians, photographers, and internet users. What, what did, what was that startup like? What, what was the- ... what was the, the gist?

[00:01:30] Luke: Yeah. The, that's, that's three different things that are kind of pummeled t- together there.

My, my, my first thing is I had a organization. I, I had a, you know, company called Rockweb, uh, that provided, uh, online spaces for bands that, that started- Oh, cool ... when... Be- before that was a thing. You know, my, my opening pitch to the bands that were on it used to be, "There's this thing called the internet."

Got it. Got it. And, and you're, you're gonna be interested in it. Uh- So it was like a

[00:02:02] John: Squarespace for musicians and photographers and,

[00:02:05] Luke: and whatnot. Um, fir- first Rockweb was for musicians, and then a few years later, uh, starting in around '99, um, I had a company based on the domain name photography.com that, that did- Oh

kind of early digital photography things. Um-

[00:02:22] John: Very cool. And- So, so a web pioneer.

[00:02:26] Luke: Yeah. Yeah. The, the internet users group was an internal thing when I worked at Cisco, and I, I was, I was Cisco's first internet webmaster. And, and so- Wow ... uh, a- again, it was... I tell you, I go back a long ways. So, so it, you know, the, it, it was another situation of there's this thing called the internet and here- Very cool

here's what it is, so.

[00:02:49] John: Very cool. I'm John Picozzi, solutions architect at EPAM, and today my cohost is Nic Laflin, founder at nLighten Development.

[00:02:57] Nic: Good morning. Happy to be here.

[00:03:00] John: And just a programming note, uh, Scott Falconer will be joining us again next week for, uh, for his show. He's actually at Acquia Engage Paris right now.

So if you happen to, uh, say, uh, see him you can say bonjour. There you go. And now to talk about our module of the week, let's turn it over to Martin Anderson-Clutz, a product marketing manager for Drupal at Acquia and a maintainer of a number of Drupal modules of his own. Martin, what do you have for us this week?

[00:03:32] Martin: Thanks, John. Have you ever wanted to use AI to suggest related content on your Drupal site? There's a module for that. It's called AI Auto Reference, and it was created in June of 2023 by Scott User, or, uh, sorry, AKA of, uh, Soapbox. It has a 1.0.0rc4 version available. It is actively maintained and has, uh, security coverage, or I suppose I should say it's opted in but needs a stable release.

It does have test coverage, and there are four open issues, one of which is a bug on Drupal.org, which officially marks the module as being in use by 19 sites. Now, AI Auto Reference works with any reference field, so it could find suitable taxonomy terms, nodes, or what have you. It does that by rendering a specified view mode, so it should work with any kind of complex layout approach you may have implemented on your site.

It will also automatically shorten your content to fit within your AI model's token window, which you can also configure. The module extends Drupal's main AI module, which means you can select which model to use, and probably means you can also use guardrails and all the other powerful features that come with that ecosystem.

AI Auto Reference comes with default prompts, but you can edit those if you really wanna make sure you're squeezing out every drop of relevance. You can also choose for which fields in each content type you want to generate suggestions, as well as whether you want the suggestions to be automatically applied or whether you want to have them manually reviewed.

Now, as mentioned on the project page, you can already have AI suggest things like tags using the AI module without this project, but this may be a better choice if you want to make sure the recommendations stick to an existing set of content. So let's talk about AI Auto Reference

[00:05:28] Stephen: Hmm.

[00:05:29] Nic: I think this is interesting, but the, the part that I'm interested in more than the auto, uh, tagging part or the auto references, I wonder what the UI looks like for the approval process, because I feel like, um, I, I feel like in this day and age, that type of workflow is gonna become more and more necessary, and coming up with a clean way and an easy way to be like, "This is what was changed, this is what we're doing," um, will become kind of paramount.

I, I, I'm curious if you had a chance to look at that.

[00:06:05] Martin: So I haven't had a chance to sort of play around with it firsthand, but my understanding is that it uses the field widget actions button that also is part of the AI ecosystem. And so you have sort of a, you know, a button that you can tie directly to a specific field, and then once it has the suggestions, it'll basically put those into a modal, kind of an overlay.

[00:06:25] Nic: And does it handle the... And the, like, it, it won't, like, re-trigger itself when it saves the, uh, the references, I, I assume?

[00:06:38] Martin: No, my understanding is it will not do that. Although I, uh, the other th- part that I thought was a little bit interesting as I was digging through it, it seems like it's really, um, because it uses the rendering of the, the sort of content to reference, it actually, as far as I know, doesn't work too well for new content.

So it's more like you would save the node, then go back into edit and say suggest the things to reference, and then at that point you'd be able to get them because it wouldn't have, like, a node ID yet, as an example.

[00:07:07] Nic: The my, my favorite, my favorite workflow, but that, that's a, that's been an on- uh, an ongoing issue with Drupal is, like, how to handle the initial insert versus an update.

I mean, we've, we've all had to work around that. I don't, I don't- I verify,

[00:07:23] Martin: yeah.

[00:07:25] John: I, uh, yeah. Uh, I'm wondering at a higher level, uh, Martin, you, you come to us with a lot of great, um, AI-focused modules. Like, how do you keep them all straight? I'm assuming you're not using all of them. Maybe you are, but, like, I'm just wondering, like, is there a, is there a method to, like, keeping them all kind of, like, organized?

Or is it just like, "Hey, let's try, try it, see what works," and, like, it's there if I need

[00:07:54] Martin: it? It ... Yeah, I mean, it's, so I guess we're really venturing off topic a little bit, but, uh, to be honest, the process is if I'm at a conference, if I'm talking to somebody, um, if I look at, uh, you know, one of the, the Drupal newsletters and I see something that looks interesting, usually the first thing I'll do is go...

W- we have a, a Talking Drupal Slack space. There's a module of the week channel in there, so I'll kind of throw it in there. That's almost my, like, my scratch pad.

[00:08:21] John: Hmm.

[00:08:22] Martin: And then I'll try and remember to actually transcribe it into the Trello board that we use to stay organized because after 90 days, everything in our Slack workspace will disappear.

Yeah. So I have to remember to do that on a timely basis. And then, you know, once, once I see what the topic is for the, the upcoming show, then I'll go in there and sort of try and look through. Usually I try and keep the more recent ones at the top, um, just because I feel like m- more recent should, should be favored.

[00:08:49] John: So- There's- Sorry, to clarify, I was, I was more asking, like, about, like, implementing these into a site. Like, is there just, like, do you have, like, just one big, like, AI mega- Recipe ... mega site that you're just, like, running all the AI

[00:09:01] Martin: on? AI mega recipe. No, that's a great question. I mean, I... As I said, I haven't used this one myself.

Um-

[00:09:09] John: Got it.

[00:09:10] Martin: Yeah. I've, I guess it would really depend on exactly which, which capabilities you need, and then sort of understanding, you know, like, of the, the different ones that can do, you know, either, uh, generating suggestions from scratch- Hmm ... or, you know, as, as this one does of, of sort of, like, picking from existing ones to, to say which ones are, are most closely associated.

You'd have to sort of fine-tune that, and then hopefully the names are- distinct enough, but I can definitely see where-

[00:09:39] John: Yeah ...

you

[00:09:40] Martin: know, without some kind of a cheat sheet, it may get confusing in the sort of back-end UI for sure.

[00:09:46] John: Yeah. We all know, we all know naming things is hard. We've all been there before.

Um- Mm-hmm ... this is actually- I know ... interesting because I know you and I had talked about kind of Tagify and using AI with the, like, Tagify field, and then having Tagify, like, style things a little bit differently as to, like, "Hey, we just suggested this, and it's a new term that we created versus a existing term."

So, um, this is a interesting, uh, this is an interesting module to kind of look at to kind of help with that, hey, select these ex- the existing terms for me as opposed to, like, me having to do it myself. So, um, could be, could be interesting. Could be interesting addition to my, to my blog. I don't know. We'll see.

Well, Martin, thank you as usual for bringing us an amazing module of the week. If folks wanted to suggest a module of the week or just catch up with you, how best could they do that?

[00:10:39] Martin: We are always happy to talk about potential candidates for module of the week in the Talking Drupal channel of Drupal Slack, or folks can reach out to me directly as mandclu on all of the, uh, Drupal and social channels, and I will also be at Ash- Asheville Drupal Camp next month, so maybe I'll bump into some listeners there.

[00:10:59] John: Awesome. Great. Thank you, Martin, and let's pass it over to Stephen to tell us a little bit about, uh, NedCamp.

[00:11:08] Stephen: Save the date for New England Drupal Camp 2026. Join us November 13th and 14th at Rhode Island College for two days of learning, connection, and community with Drupal professionals from across the region and beyond.

This year's conference features our popular Higher Ed Summit, a full lineup of speaker sessions, and both half and full-day training opportunities designed to help you sharpen your skills, whether you're new to Drupal or a seasoned expert. And of course, it wouldn't be NedCamp without time to reconnect with friends and make new connections at our community social.

Visit nedcamp.org for more details, and mark your calendars for November 13th and 14th at Rhode Island College. We can't wait to see you at NedCamp 2026.

[00:11:56] John: Thanks, Stephen. I, I couldn't agree more. I'm really excited, uh, for, uh, for this year's NedCamp, and, uh, we'll be, we'll be getting, uh, session submissions opened, uh, hopefully the end of this week or beginning of next week, and then we'll be announcing our, um, very cool keynote on, uh, uh, for the Saturday. So, um, just as a, a teaser for anybody that's listening and, and that, that's interested, uh, we have, uh, an elder statesman of the web coming to, uh, to talk ab- to us about, uh, what he's seeing f- for the web in 2026 and beyond.

So, um, it should be, should be pretty interesting. All right, Luke. So let's, let's fast-forward to, fast-forward to the future here, and we're gonna, we're gonna talk a little bit about, uh, AI, but, but more specifically agents. Um, so could you just describe, like, w- what is the agent management system?

[00:13:00] Luke: I'll try.

That's my job here. Uh, it, it's, it's basically a methodology that I've been putting together and, um, kind of half stumbled into it. Uh, I'll, I'll, I'll give you sort of the, um, the, the birth story is, uh, in, in the fall of... What were we in? 2024, I, I guess. Um, was it 2025? The years all blend together. Um, might've been 2025.

Might've been that recently. That, uh, Anthropic, uh, did a promo where they offered, uh, $250, I think it was, of credits that had to be used in a two-week period of time to, uh, to Claude Pro users, which I was on a $20 a month plan.

[00:14:02] John: Mm-hmm.

[00:14:02] Luke: So, so I went from the stage of, of, okay, you know, I've been, you know, very carefully metering my AI usage to try not to, to run out to like, holy cow, you know, there's, there's enough stuff.

So I, I tried to use it all. I tried to like, okay, what, what's the most ridiculous thing I can think of that, uh, that I could ambitiously try to get the models to, to do? Um, so I, I, I work a lot with the Backdrop project, which, uh, I'm sure you guys are- Mm ... are, are familiar with. Um, so I decided, well, what the heck?

Why, why don't I, why don't I try to create a system that would let Backdrop use any WordPress theme as its theme, uh, by incorporating a complete running copy of WordPress inside of Backdrop, inside of Backdrop theme, which is, you know, a ridiculous, au- audacious idea. Um, and so, so I just, you know, threw Claude at it, and, and it, like, almost solved it right away.

But then I hit context limits, and it ran out of things, and I kept restarting it, and I ended up forking. And, and I discovered, which, you know, a lot of us have along the way, that, um- AI agent tools, um, can solve problems really quickly, like astonishingly quickly, and then they forget what they just did, and it's hard to follow along with what they're doing.

And it, it turns out that, that finding solutions is, at least for me, a, wasn't, wasn't an unsolved problem anymore. The, the unsolved problem was retaining those solutions. So I, I kind of started looking for, uh, ways to do this. Now, now I've been, you know, I do some programming, but, but I- I'm mostly a, a programmer manager.

I, I've run Scrum teams since, you know, 2012 I think was probably my, the first time that, that I was a Scrum master for a thing. So, so I'm used to writing up specs and going it out to developers to go off and do the work and then evaluating their results. And, and I'm, I'm a big fan of Scrum. Uh, I, I like sprints and stories and, and all those pieces.

Uh, I, I find that the So bureaucratic overhead o-of Scrum that seems annoying sometimes when you're in it is, is a very good at providing kind of like a ratchet. Like if you, if you get something solved in, in Scrum, especially if you're storing your results in, you know, a tracking system, you know, like Git, once you achieve something, it's locked in.

You've, you've got that. Whatever happens next, you know, the worst that happens is you fall back to that last like ratcheted point. Um, so, so I tried, um, you know, firing up Jira, you know, personal Jira account, and putting together like another project and, you know, I've... E-each one of these, you know, it was like, "Okay, it's a new weekend, what crazy, ridiculous thing that I've been dreaming of trying to do for decades can I now just do in a few hours?"

And, you know, so I'd, I'd plug each of these program, you know, each of these projects into this. And, and I found that, that actually Jira seemed to help, you know. It, it seemed to, you know, with me telling it, um, you know, managing the communication between things. Anyway, so fast-forward a little bit past that.

Trying SpecKitty actually was an important one. I, I need to throw that in 'cause Robert Douglas is doing kind of brilliant work with this very, you know, organized, uh, um, harness called SpecKitty and, and that, that tipped me off to a lot of the things like skills and stuff like that, that are, are built into Claude.

Very useful stuff. Um, and it actually builds a Kanban board, which is, you know, I was like, "Hey, that's great." Um, anyway, so, so where, where we enter the thing that I kind of came up with, Agent Management System, uh, was, was just this last January. Uh, and, and it's funny because I kind of coined a term that I now see all over the place.

So like I wasn't the only one to, to call things handoffs, and I, I actually see that all the time now in AI world. It, it wasn't a thing when I started. So I might have to find a new name for this, uh, 'cause, 'cause I, I might have lost that particular, uh, namespace. Um, but, but I came up with, with literally a one-page project.

The, the initial handoff protocol that, that's in my, uh, GitHub repo, uh, you know, github/celliar Uh, started out as just the README. Like the README was the entire thing. And the, the, the way you would use the system is you would download the README to some place in your project, tell your agent, "Read this and follow this," and it would do it.

And, and basically there were, there, there, there were two pieces. There's two folders full of files. So it's all file system based. It uses the file system, you know, markdown files as the, um, a- as the organizing, um, principle. Okay. There, there's a handoffs directory and a docs directory. And so, so the, the handoffs, uh, describes a, uh, a file name convention that, that tells the, uh, the agents when they finish, like a chunk of work.

And this is typically like commit level, right? Like, like when you're, when you've, when you've solved some piece and you're ready to do a commit, like as it writes the commit, it, it writes a dated signed handoff file with, you know, the, the, the date so that it sorts and its, you know, model name. You know, I'm, you know, Claude Sonnet 4.3 or whatever, like did it.

And, and it's just like a little outline that it says what you should put in the, the handoff. Says, you know, and, and it sounds a lot like, like the, the scrum standup things like, like what you did- Okay ... what you tried, what didn't work, what are the open questions, what artifacts did you change? Do- stuff like that.

And I kind of thought when I was doing this that this was like the, like first step in what was gonna be a long process to put together maybe some big system. But here's the thing. Here's like the weird thing I've discovered, is this incredibly vague spec, basically. The, the agents are remarkably consistent in correctly interpreting them.

Like, like typically if, if I, if I tell it to follow this stuff- There'll be variances in, in what they do for it, but they, they come up with like enough, close enough that, that it, that it works. And, and if there's other handoffs already there, they will read those and incorporate that style, and th- they're very good at just kind of, you know, picking up what the crowd's doing and, and moving along with that.

Um, so, so just this one thing I, I found just incredibly, incredibly useful.

[00:21:41] Nic: So, uh, so help me, uh, help me put a finer point on that then. So it, it... Or, or correct me if I'm wrong. So you're saying, so the agent management system then is, um, a little bit different than kind of what most people are doing with skills and things.

It's more of like a here's how you identify what you did situation, so you can more easily recapture that context in future iterations?

[00:22:10] Luke: Yeah.

[00:22:10] Nic: Or is it something different?

[00:22:12] Luke: Well, well, I... It depends on, on what you're, you're meaning by you in there. 'Cause, 'cause the, the, the markdown files, the, the handoffs that, that it magically creates, um, in, you know, all sorts of different contexts.

Like, it's really funny, like you, like- You know, I've, I've done this, you know, dozens of times in different contexts and, and it just seems to work. Uh, and, and it's typically, it's typically the agents that are reading the other agents' handoff files. So, you know, th- that's part of the protocol is, you know, when you get in, read the last few handoffs to, to orient yourself.

That's, that's the first instruction in the, the handoff protocol. Uh, and so typically they're passing messages to yourself, but they're, they're readable. It's fully like, you know... It, it's a, it's a markdown thing, so you can double-click and like any of the handoffs read like, you know, good solid documentation.

They're, they're, they're little one-page, two-page, three-page memos about the, the state of the project. They're, they're the kind of thing you wish your programmers would do, right? Like at the end of the day, if they would send you an email describing what they did that day, it looks like this. Uh, so, so in terms of sort of the group thing, this...

It, it's effective even if the group is just you and one agent. Like, like, that's enough to get value out of it, uh, because between you and a robot, you can keep track of what you're doing. You can, you know... You know how it is. Like you're, you're working and then like you come back the next day- Yeah ... and you're like, "Oh."

Or the end of the weekend, like, "What was I doing?" Uh, whereas in this case- Well, I- ... if there's... If, if you're running this, y- you can ask the robot, like, "What were we doing?" And it will explain- Yeah. ... in your language. Catch up.

[00:23:54] Nic: I, I'm, I'm curious then. Do you have, um... So I, I see your GitHub repo has a bunch of stuff around this, a bunch of protocols.

But do you have- Yeah, yeah ... examples of what the output looks like as well, or are those kind of on your, on your local drive? 'Cause I think it'd be useful to- Yeah ... so for example, I, I have a client that's been using, um, Claude a fair bit to, um, put together, you know, different proof of concepts, different feature requests, that kind of thing.

And, and we have a... They, they have a similar protocol where they output like, "Hey, this, this merge request does XYZ. Here are the features we're adding, the bugs we're fixing," et cetera. And one, one thing that we've run into is that... I mean, this is just a general AI issue, is that- Yeah ... they're, they're just so verbose, right?

They're so verbose that it's hard to re... I mean, they... When you read them, they generally look correct, right? They, they generally have, um, pieces in there. They'll have pieces that you forgot. "Oh yeah, I forgot we tweaked that. We f..." Like, they, they capture that stuff, but they're so verbose that they're hard to, um, they're hard to grok, you know?

And you can use- Yeah, yeah ... you know, you can use AI to, to parse it. But, uh, I guess my question is if you're using AI to summarize an AI-generated thing, why not just store the summary, right? Um, as well so that you can, you can see that too. But I, I'd love to compare this with that output to see if it provides a more succinct kind of- explanation of where, where you are?

'Cause I, I find that that's the hardest part is just like- Yeah ... really quickly catching up as a, as a human on where you are and what's been done. Yeah. Um, do, do you have those examples?

[00:25:38] Luke: I, I do. I, and that, that's a- that's actually part of it. They, that was sort of the, the beginning stage was, was handoffs.

And, and note, by the way, that I only really talked about half of the files thing, 'cause there's the handoffs directory, which includes- Okay ... just these daily memo things. But there's also a docs directory, which is for longer term things. And, and again, that's kind of dynamically decided. Uh, it, uh, uh, I guess actually an, an important part of this that, that's really fundamental is it- It's excessively human in the loop.

The, this is the opposite of agentic programming. So you, you don't like write a nine-page prompt and send it off and have it do it. You are there every step of the way, hearing every message. As they're talking to each other, you're part of it. At least I'm part of it. So, so you... The, the idea is that you're there feeling the pulse of the, the changes.

Uh, you can- Right ... if, if it ever g- You know, you're, you're not gonna come back and find that it's like gone off for hours doing something crazy. Like, if it does something crazy, you know right away. You, you, you see it do the thing, and you're like, "That's ridiculous. Why on earth would you do that? No, do it again."

Like, like, it, it, it involves you each step of the way. Like, like literally, I, I mean, sometimes things will run for like a minute or two, or if... Like, if it, if it ever runs for like five minutes without talking, I'll, I'll stop and say, "What are you doing? Explain yourself. What, what are, what are you caught on?"

'Cause like it's designed to be very, very incremental and bite-size. Uh-

[00:27:18] Nic: Mm-hmm ...

[00:27:18] Luke: and, and, and so there, there's... On, on my repo, for example, you, you're asking if there's, if there's sort of examples of this. Uh, I have two projects that, that I did on one Hawaiian vacation, uh, that were, um, extensions to DDEV. Uh, o- one of the like outrageous things I've wanted for years that, that I did is, is I, I wrote a Xdebug, um, plugin for DDEV where you, you can fire it up and it, you get an Xdebug interface in the terminal window- Mm

uh, for, for whatever DDEV project you're on. Um, which-

[00:27:59] John: Oh, really? ...

[00:28:00] Luke: I was super proud of. I very excitedly showed it to Randy Fay at DrupalCon, and he was like, "What's that good for?" That's why Randy's the nicest guy in the world. He like apologized excessively for like kind of blowing me off with that, but, uh, it, it just kind of shows how, how you gotta- Right

you gotta kind of market test things to do it. But the, the key thing is, is that in those repos actually, I, I've got, uh, a list of the stuff that it kicked off. Like, like besides even just, you know, the handoffs and the, the docs, um, since I was, you know, writing a project in Go. I've never seen Go. I, I didn't know how to do this.

So, so I created a, a folder called Learnings, and along the way, as it wrote all these Go things, it explains all the Go techniques that it uses in terms of a, uh, Drupal PHP programmer would understand. So it translates, you know, Go concepts to what PHP programmers use. Um, you know, there's Go, Go version of Composer, for example.

[00:29:08] John: Going back to, going back to the original question of like what, what is the agent management system? So it's, it's, it's less of like a, a computer program and more of like a library to help ma- your agents better communicate with each other, and to, to be, to be kind of succinct on the status of your project.

Is that a, is that a fair as- assessment?

[00:29:35] Luke: Yeah, yeah, yeah, it's getting there. I, I, I would say toolkit, right? Right. Like, like, like sort of, sort of like if you're using Unix, you, you might use Grep for something. You, you might use Set or Arc. Right. You use whatever pieces are, are appropriate.

[00:29:48] John: So it's like, it's, it's like a knowledge, it's like a knowledge base of your project for agents to be able to, uh, be able to kind of work with each other more efficiently, right?

Um-

[00:29:59] Luke: Yeah. It, it, it, it's curated though. Um, I, I actually wanna throw in a mention... B- before I did all this stuff, I, I, I have a, a, a site called Simplify Drupal, uh, simplifydrupal.com, which it's been stale for a little while. It's been, been hanging out there for, you know, I, I kind of spewed out the, the ideas that, that I wanted and then, you know, I've left it there.

I'll, I'll get back to it at some point. But, uh, but, but it... I, I am, I'm eternally pushing back against complexity. I, I, e- every time, you know, I've, I've been in this game long enough that, that I know that just complexity grows and grows and grows, unless you're like- actively, actively fighting it, y- you just always get overwhelmed at some point.

Uh, so, so bringing that forward towards the AI stuff, my, my idea is to, like I fought the instinct to, to make it, to make the original handoff protocol more than one file for a while. Finally, I decided nah, it's just too, you know... I, I needed a readme that people see when they go to re- GitHub that's like human-directed, and then like a slightly less readable, like, set of instructions for the robot.

So I boiled it up to two, two files.

[00:31:12] John: So I'm interested in, like, you have this structure of the AMS, right- Mm-hmm ... with, with the, the handoffs and the documentation and the learnings and like, is the AI, a- are the agents, um, actively like writing to those directories and providing that backlog of information or is that something where you're like, the agent does something, it provides you with an output, and then you take the learning and then you put it into that folder?

Like, are you curating that, that list of things-

[00:31:46] Luke: No ... that, that- You're, you're, you're not... I mean, that, that, that, that was the initial thing, and, and they... That's, that's why I find Claude Code and Codex, like, way more useful than chat interfaces 'cause-

[00:31:57] John: Hmm ... '

[00:31:58] Luke: cause having, having them able to directly read and write the files- Mm-hmm

removes just a huge amount of friction. Uh, it makes it so much easier. Uh, so, so I'm only involved really in that they'll tell me. Like, like the... I'll, I'll kind of fast-forward c- 'cause there's, uh, there's other pieces, but, but the, the working system that, that, that I use in production all the time now is, is I'll have- Typically three to four, um, agent threads defined.

And, and I give them names and I give them roles. Uh, so, so, you know, my favorite is generally I have a librarian who is responsible for the documentation, is responsible for the specs and, you know, communicating to the humans and kinda maintaining that. Um, and I'll have a separate coder, um, agent. And, and it saves money because they can be different models.

So the, the librarian is often in like a fancy expensive model as in Opus or I was all ready to try Fable this last weekend. Uh, but the coder is generally Sonnet, uh, a much cheaper thing. And I use Haiku for things all the time, which is an even cheaper model. So, so I, I had a, I, I had one month of a Claude Max, the $100 a month, um, uh, subscription, but, but I'm currently on a pro plan, just a $20 a month plan.

And I, I do it as kind of a discipline, right? 'Cause, 'cause it's easy to just throw power at things. But, but if you... It, it's kind of, um, uh, what do they call in music? Uh, you know, creative, uh, constraints. You know, kn- knowing, knowing that I've got, you know, a certain amount of Claude token usage in a five-hour period, it, it just slows me down a little bit, makes me think, makes me position things in a way.

And, you know, I'm, I'm fully ready, you know, this is important enough to me that, that if I was limited, if I was stopped, I would, you know, it's worth five bucks a day for me to have Claude Max. But I haven't needed it. Like, I'm, my credit card's ready. I'm ready to like- Put it up. But, but I get, I get all my stuff done on Pro

[00:34:19] Nic: Uh, so I'm, I'm curious then if you could talk about how you decide which model to use for which task then, because I, I think most people just, that are running agent systems and putting together these types of scenarios are just, they're all in.

They're paying for the max, and they're just- Yeah, yeah, yeah ... throwing everything that they can at it. And- Yeah ... when, when the, when the true bill comes due, they'll be sorely they'll be in trouble. Um, uh, how, how are you deciding... Like how did you decide, for example, that the librarian needs the most and the coder needs second tier?

Um, uh, uh-

[00:34:53] Luke: Yeah, yeah.

[00:34:54] Nic: Well, well- Are you running tests to get on, on different models to see what the output's like? Are you just using your gut? Are you, um, how are you, how are you- I'll

[00:35:02] Luke: answer the question if you give me a chance.

[00:35:05] Nic: Yeah.

[00:35:06] Luke: I, I mean, my, my, my first step is, is I ask them. Uh, and, and I find that super effective.

Like, like, like, like I'll start generally with, with the first plan. I'll, I'll use a heavy model if, if I can afford it. I actually don't really even need it for most things, but you know, if you have it, I'll, I'll say, "Okay, Opus, you know, let- I'm... let's do this big, ambitious thing. How would you break this down into pieces?"

And then I just immediately ask them. It's like, "So what is, what is the least expensive model that will do this?" And like most coding tasks, uh, it says, you know, Sonnet handles it fine. That, you know, Opus is overkill, doesn't need to do all that thinking, doesn't need to spend all the tokens. Um, I, I use, I use Haiku for things.

I mean, people complain for... Like, like, like people talk about using, you know, caveman mode to try to save tokens by, by speaking in, you know, abbreviated things. Yeah. And I mean, that seems, that seems silly to me. I, I just fire up a Haiku session, which you could run all day for hardly anything, and, you know, ask it back and forth questions.

So, so like, like a typical spec, a typical process for me with these things is, is I'll say to Opus, it's like- Write out a questionnaire about this topic for Haiku to feed to me. And it writes out a little markdown file and a prompt with the path to the file just written. I just go to my Haiku window and just literally just type in that path that I just copied from the Opus window.

And Haiku pops up, it's like, "Okay, I'm here to, you know, give you an interview questions. First question is, you know, how about this? What do you do this?" And I type out the answer. "All right, how about this?" And, and so it doesn't really even process that, it just records it. But it, it's smart. Like, it's still an LLM, so I can, like, adjust things along the way.

It's, it's not, it's not like a voice recorder. I mean, it, it, it's an intelligent thing that's interviewing me, but it's a cheap one. And then when it's done, I go back and have, um, you know, Opus interpret this, figure out sort of the strategy, and then it, you know, writes out a very detailed spec. You know, the kinda like one-shot stuff that we talk about.

But I don't, I don't spend three hours writing a, you know, nine-page prompt. I, I answer the questions and let Opus write it, and then it feeds it to, um- It feeds it by giving me the file name of the spec file, and, and I'll, I'll open up the spec and read it to make sure it's, it's not, like, draining my bank account or something, right?

But if it... You know, generally it's straightforward. You know, it's English. Yeah. If it's not, I'll send it back and make it rewrite it in English, right? Like, it does what I tell it to. Uh, and then I, I pass it to the engineer, and he goes, "Oh, I got the spec. Here's it." It's... The engineer might ask some questions about it, might make some changes, might say, you know, trade-offs, gives me little polls, like, you know, "Do, do you want, do you want it like this or that?"

Um,

[00:38:09] Nic: so- I just wanna, I just wanna call out one, one of our listeners said that he enjoy... Kaiser said he enjoys the caveman mode just because- ... it makes it more fun, so-

[00:38:17] Luke: I know Justin likes that.

[00:38:21] Nic: Um, I, I, I wanted to kinda get back to the, the global AMS system a little bit. Uh, AMS system, AMS system system. Uh- ... it, is this Drupal specific?

I mean, I think you mentioned you used it to write the DDev add-on, so it sounds like it, uh- Yeah ... how you're using it, you're using it on more than just Drupal projects and using- Yeah ... a backdrop.

[00:38:43] Luke: Not, not at all Drupal specific. Um, but you know, it's, its roots are coming from me, and I'm a Drupal guy. So, so, like, like, it is thoroughly tested, um, on the tools that I use, you know, writing Drupal with Claude on a Mac in DDev.

Uh, like, it, you know, with Jira as a back end and things like that, it, you know, it, it, it's tuned. It works really well with the tools that I use 'cause I've tuned it to really work, work well with the DDev, but, but it's not specifically, um, tied to them. Uh, I, I use the same things actually for, for non-computer projects- Mm

that, that come up. It, you know... I'll, uh... I'm, uh, I'm, I'm trying to think of a, of a time I've used where it's needed more than one, uh... I li- kind of, kind of marketing things, for example. You know, I, I show it my domain portfolio and, and kind of can bounce back and forth between things. That, tho- those are actually, th- those aren't really good examples of this 'cause, 'cause tho- those are just talking with an agent.

Um-

[00:39:49] John: I wonder, I wonder, so y- you said you, you started it with Claude. So it's not Drupal specific. Obviously we could use it for, for other, other types of projects, other types of tech, which is awesome. But, like- Yeah ... you, you said you started it with Claude. Is it AI specific? Like, could I tie it into, you know, ChatGPT or, or OpenAI or, or one of the other, you know, uh, big AI providers as opposed to Claude or is it- Oh,

[00:40:14] Luke: yeah, totally.

[00:40:15] John: Okay.

[00:40:15] Luke: No. No, it wor- works, works fine. And, and in fact, you know, I re- as I say, I like to find the limits of, of where the, these things go. Sure. So, so one of the, one of the two, um- Uh, DDEV extensions, like Xdebug was one. And I also did, did another proof of concept where it, it puts an interface to Drush. So, so for that, um, I specifically, uh, told it to consider, um, other models.

And, uh, and so, so, you know, part of... I- th- there's actually other aspects of AMS that, that I should probably ge- get to. I, I mean, I mean, a, a big part of it is, is things for using Scrum with the models for, for dividing things into stories and sprints and things like that.

[00:41:09] Nic: Okay.

[00:41:09] Luke: Uh, and, and so, so for the, um, for the Drush DDEV project, um, each, each sprint w- was kind of defined as when the U- UI changes, right?

Like, like the, the, the sprint goal for the initial UI, uh, w- for the initial one was, was to get it to, to, to print a screen, you know, in the, uh, in the DDEV container, you know, handle all the cursor stuff. And, and that broke down into like three or four stories. And, and the initial, you know, project manager Opus thing would suggest which models could take on those stories.

And so it would, it would suggest Anthropic models like, you know, Sonnet or, or, um- Haiku could handle this. But it would also say, you know, this is good for Codex. Uh, we can use the, uh, Gemini CLI for this. We can use Composer CLI, which is Cursor's internal model. Um, so, so it seems, seems to try, it seems to work, like, like there's nothing super specific, um, anthropic about it.

And I find actually that the other models are able to handle commands I thought were anthropic only. Uh, like, like I have a skill, for example, for the handoff protocol thing, that if I just in a, in a Cloud Code session, you know, a new project type/handoff, it goes and downloads handoff and reads it and writes the handoff directories and files and like, you know, kicks off the system to start to do that.

And to my astonishment, it works in other models too. It works in Composer, it works in Codex. I don't know how, but like, it just does.

[00:43:01] John: Is that, is that concerning? Is that concerning to you a little bit? Like, you're like- Yeah ... I, I don't know how this is happening, but it's happening and it seems to be happening right, so I'm just gonna kinda roll with it.

[00:43:12] Luke: I, it, it is, but, but, but here, here's the thing. Uh, I am building stuff all along to involve... The thing I'm trying to maximize is the value of the human in this. Um, so as the, as the robots tech, you know, capabilities change-

[00:43:32] Nic: Hmm ...

[00:43:33] Luke: humans, though, we have a certain amount, uh, it- there's, there's a term for it that, that I found.

This is what I've, was explaining with Simplify Drupal, is, um, cognitive overload.

[00:43:44] John: Mm-hmm.

[00:43:45] Luke: There, there's only a, there's only a certain number of things that you can keep in your mind at one time before it just starts to bleed out and, like, you lose track of what's going on. Uh, so, so that's, that's the main thing that I'm trying to maximize, and I'm trying to keep in mind like, like what's the maximum human cognitive overload.

And I try not to like ... I, I mean, this is my complaint with a lot of computer stuff, is, is people tend to complicate things to fill their, fill as much as they can think about. Um, and you can see in Drupal, for example, right? Like, like around 2014 Drupal expanded to the point where one person couldn't do it all well.

You had front end and back end, and like the front end people, you know, created a whole more complicated thing than they need to be. Like, I hate SaaS, for example. Like, like, you know, okay, it does things, but like, it's just another like 100 commands you have to memorize in order to get something done, right?

And, you know, whole tool chain thing. Like if that's all you're doing all day, that's fine, but if you're doing that and like 10 other things, if you're also a lawyer- Mm-hmm ... or something, right? Like if you've got something else that's competing for space in your mind, it's a lot of stuff. And if you can get the, get the job done with a smaller budget of neurons, you know, it's worthwhile.

And, and that's, that's kinda what I'm trying to maximize with these things. I, I want you to be able to use, use the tools, but, but not be a full-time job, right? They... It should be so easy that you're like, "Oh, that was easy." Like- So, so- Like, it shouldn't be something you have to learn ...

[00:45:20] John: so like, how do you toe that line, right?

Because a lot of like, uh, and I've heard you mention it twice now, of like human in the loop being like, you know, human watching what's going on. Yeah. But a lot of like what AI agents are great at is like, go do this thing, and then go do it, and then you come back and you tell me like, "Oh, here, I did it.

How did I do?" Right? Yeah. So like, it's less about the human, like, watching it and, and, and monitoring what they're doing, right? It's more about like, "Go do this, come back and give it to me." So like how do you toe that line to make sure that, you know, you're, I don't know, not overly involved where you're losing the benefit of the, of the agent itself?

[00:46:02] Luke: Yeah. Well, it, it's, it, it's a, it's a line that, that I toe. I mean, I mean, I, I adjust the settings, for example, that talk about like confirm every action, write things, do things auto. Uh, I, I tend to- alter that as the project goes on. I start with it meaning being like, ask me for everything until I make sure, like, like once it's kind of gone through and been like, okay, you know, I did these parts, I, I trust it more and more and tend to put it in auto, give it more and more permissions.

Uh, but, but it- it's like any kind of management thing, right? You know, you, uh, a new programmer comes on board and you give them an easy project and you see how they do.

[00:46:46] Nic: So if I, if I'm interested in testing this out myself-

[00:46:49] Luke: Yeah ...

[00:46:50] Nic: how easy is it to set up? And, and what does that process look like then?

[00:46:54] Luke: Yeah.

Well, uh, the, the easiest thing... And, uh, the... One of the reasons I came up with AMS is 'cause I needed like a easier brand name for this. Like, like literally the answer to that is, is to go to my GitHub fol- GitHub repo and go to AMS, and that's basically a directory to these other things. Like, like the first thing is the Handoff Protocol, which I might have to rename 'cause Handoff's become a word.

Maybe I'm calling it Memo Protocol. I might, I might have to come up with another name for that just 'cause Handoff seems to have like gotten out of my reach. But for now, it's still called Handoff Protocol. Uh, and, you know, you, you can go to the repo, it'll tell you how, you know, you can just read it, and the README will tell you what to do, and you can handle it, uh, easily.

I, I don't actually, you know, it came out more insulting there than I intended to. Like, it's, it's easy. Anyone can do it. Uh, and, but, but there, there's a skill, uh, a... That, that's in there. The, the handout... Let me see. Actually, I can look into my repo here for a second here. Uh, I think it's called Handoff Skill, maybe.

Um, I have to make this window a little bigger. There we go

[00:48:10] Nic: So, so-

[00:48:10] Luke: Anyway, uh, yeah. Oh,

[00:48:13] Nic: yeah Yeah. So, so, so is it as simple as just cloning this down into whatever code space that I'm gonna be working on with Claude or- Yeah, yeah, yeah ... and then doing the commands?

[00:48:22] Luke: Yeah, yeah. Ba- basically, it's, it's a single file. It's an agent.markdown file that, that's the two-page description for, for the Handoff thing, for like the basic thing.

And all my other tools assume this is in place. Like, like this is, this is the piece that they all use as their communications kind of substrate. Uh, but it, it's just an agent MD file, and it will create, with your permission, um, folders. Uh, by default, they're just called Handoff and Doc. You can put them in an AMS folder if you kind of want them out of the way.

You can put them in a .AMS folder if you want them really out of the way. Like, like you... It, it's, it's customizable simply by talking to the agent. Like, if there's anything you don't like about it, you can tell it to change it, and, and it's, there, it, it's very easily customizable. But you don't, you don't have to know that.

You can just start running it, and, you know, you don't have to have an opinion that you don't have yet, right? Like, like you can use it and decide that you don't like something and then change it, but yeah. You can-

[00:49:27] John: How does that, how does that impact future updates, right? So like, if I start AMS and I put it into a directory, you know, wherever it is, fine.

But then I ask my a- my, my, you know, my AI buddy, whichever one it might be, right? To go- Yeah ... and start, start kind of ma- manipulating, changing, updating it. Like, how do I- Yeah ... then get, you know, if there's an upstream change, how do I, how do I pull that in?

[00:49:56] Luke: Yeah, yeah, yeah. Well, it, it's all, it's all funnels down on your local system to that one file, to agent.md.

Hmm. And, and so, so it never leaves that ground of, like the first thing it does is look at at agent MD. So, so if, as I said, you know, if you want to customize it yourself, you tell it and it will rewrite agent MD for you. If there's something upstream, you could literally tell it, you know, fetch the new version from GitHub and, you know, incorporate its changes into, into the thing.

I... The way I would do it, if I was doing that, is I would say, "Fetch the new changes. Tell me what those changes are. Ask me if I want them." And then, you know, it would answer those questions. Like, like, like I don't- I, I really don't like computers doing things that I don't know what they're doing. So, so the, the, the gate is, is definitely my understanding of what's going on.

If, if it exceeds that, I tell it to stop and slow down until I am comfortable I know what's going on.

[00:51:09] John: Yeah. That makes sense. So, uh, this is all, and maybe we should have led with this, but this is all based on a, a, a talk that you did at Stanford WebCamp called Agile For Agents- Mm-hmm ... Managing Robots The Way We Manage Humans, right?

[00:51:25] Luke: Yeah.

[00:51:25] John: Yeah, yeah.

[00:51:26] Luke: Um,

[00:51:26] John: which, well, I mean, if you th- if you think about it in that, that way, it makes, it makes a lot of sense that this is a, a framework for managing your agents. And if you look at your repo, like you have a great visual there of all the different agents that you kinda get out of the box here.

Um, they- Yeah ... even have names, which is, which is, um, is, is, is, uh, you know, lack of better word, cute, but also makes- Yeah ... them very mu- much more approachable, right? Um, but you talk about this idea of persistent memory. Um- Yeah ... and I'm wondering, like, what does that, what does that look like in practice? Is that simply just, like, the, you know, the creation of these handoff files and the creation of your agents.md kind of just, like, allowing for, for that ongoing memory for the project that you're working on?

[00:52:15] Luke: Yeah, yeah, yeah. Well, well, a, you know, there, there's a current implementation and sort of the overall picture. Uh, the, the, the idea of sort of the persistent memory- Is that, is the, the LLMs do an amazing job of, like, building these ephemerables, ephemeral structures in their context. They, they kind of remember stuff.

But, but it's got a limit. It hits a limit, and then it's very unreliable when it hits that limit. Like, it starts to forget things, it starts to hallucinate stuff. So, so it's, it's a very scary world to rely on that, 'cause you don't know what you're gonna get. So, so the idea is to, like, cash out and take your profits, and like, like when you've figured out a problem, when you've figured out something, get it out of the context.

Get it so that it writes it somewhere. And, and th- these are all, like, various versions of somewhere to write it. Um, and I mean, I see all over the place on, on the... You know, people have come up with their own versions of this in all sorts of ways. So, so what I'm kind of offering here is, you know, my simplified thing.

It's like, you know, you, you don't need, you know.

[00:53:29] John: So

[00:53:29] Luke: 300, you don't have to memorize all, like, hundred Claude commands.

[00:53:35] John: Right. But I mean, so like, I, I guess the way I'm looking at it, and maybe I'm, maybe I'm looking at it wrong, maybe I'm looking at it right. But like, the way I look at it is once you get to kind of a decision with your agents or an output from your agents, right- Yeah

you're saying, "Hey, agent, write that to the handoff directory, write it to the library, write it to the documents, write it to wherever it's supposed to go," right? And it's kind of like- Write it to

[00:53:57] Luke: Jira is actually... The, the Jira MC- MCP interface is super handy.

[00:54:03] John: I, I haven't- Okay. Or, or even better, Confluence, right?

So like, I, I just- Yeah ... see it as like a way that this thing could say, like, update your decision log, for example. Like, "Hey, we just made a decision. Let's update the decision log," and then that way it's, you know, ever present for both people, uh, actual people on the project, and then agents that come to the project later, right?

[00:54:23] Luke: Yeah. Oh, yeah, yeah. Uh, and it, it's super handy for people. I, I mean, that's, that's the funny thing about it, is like all this stuff that's, that's like almost like machine code to them, li- like, it, it, to the agents. But, but it's just insanely readable. Like, at any point you can pick up any of these artifacts and, and just understand it right away.

Like, it's all, you know, well-written English. I, I-

[00:54:49] Nic: Do you... I, I mean, I- I- I guess, I guess the question then is, going back to something I asked earlier, 'cause I find, um- I, I find that it takes... W- when I, when I do try to understand these type of decision documents that are generated, right? 'Cause w- Yeah ... like I said, we have a client doing it.

I find it takes a significant amount of effort. Yeah. Like, there's no, there's no like, oh, you just, you just glance at it and you understand what's going on. 'Cause you, like you still have to like... It, it's just like any... A- and that's true of human-written documentation, right? Yeah. Like, you can't just scan it and be like, "Okay, I know where I am."

You, you need to actually process it. Um, y- I, I'm curious if you've put in any, any thought or documentation to like how, how you transfer... Uh, almost how you, you build up your own context window on these projects again. 'Cause I imagine some of them you work on for a couple weeks, and you set them aside, and you come back a month later or two months later.

Have you done any work, a- any thinking on how, how to track ways to make that more efficient, right? Or, or ways to measure the efficacy? 'Cause like you said, y- your, your goal is to make the, um, uh, used a different term, but make the burnout portion of it, the cognitive overload portion of it better, right?

[00:56:09] Luke: Yeah. Yeah, yeah. So- To, to avoid cognitive overload-

[00:56:12] Nic: Yeah ... on the humans who are managing

[00:56:13] Luke: it. So, so how are you,

[00:56:14] Nic: how are you tracking cognitive overload in all these processes? 'Cause I think that's the most important thing- Yeah ... brought up, right?

[00:56:23] Luke: I, I, I mean, I mean, you, you just said it actually. Uh, so, so in, in my case, if, if I'm in your situation and it has, you know, written like a summary document, and, you know, I open it up and I'm struggling to understand it, I, I don't beat myself up and say, you know, "I'm a bad human for not being able to understand this."

I, I beat up the robots. I say, "I can't understand this crap. You know, explain this to me. Draw a diagram. Make it shorter. You know, I, I don't want 500 words here. I want 200 words." You know, like it... Demand what you need. They work for you.

[00:57:00] Nic: Mm.

[00:57:01] Luke: You don't work for them. Mm.

[00:57:03] John: It's interesting that you say that, 'cause I, um, I'm gonna misquote it-

so I'm gonna try to go find it really quick. But I saw this, I saw this interesting, um, post on LinkedIn, um, from, uh, Jim Shaw, who, um, uh, is formerly of Acquia. Um, but he was saying... What was he saying? Hold on a second. Let me see if I can find it really quick, 'cause it, it was, it was basically like, you know, AI is like an, an inexperienced employee as opposed to like this like high level like superhuman thing that everybody thinks it is, right?

Yeah. Yeah. It, it, it often, often will get things more wrong than right, and need a, and, and need a, a fair amount of hand-holding. Um- So I thought that that was interesting. Yeah, he, he says it's the unreliable employee, right? Um- Yeah ... and I mean, I think that that's, that's kind of, that's kind of, uh- ... an interesting way of looking at it.

And, uh, you know, a- and I think that you're, you're... You know, the, the way, the way that you stated it is kinda right. Like, the, these agents are kind of working, working for you and following your direction. So if, if they're doing something that, that you're not, you're not in love with, right? You should definitely, you should definitely tell them, tell them that.

Um, I mean, I think if, if Scott were here, and I, I don't wanna, I don't wanna put words in Scott's mouth or pretend- ... pretend like I know what he would be thinking. But I think he would probably- Yeah ... he would probably err on the side of, like, you know, y- y- you can, you can kinda guide them, right, uh, to, to greatness.

But they're, they're, they may not get there on their own, right?

[00:58:52] Luke: Yeah. Well, well, the, the thing that I realize all the time when I use this is, is because, 'cause I do personify. I give them names. Um- Mm-hmm.

[00:59:00] John: Right ...

[00:59:00] Luke: if, uh, the common... Like, like, like, I try giving them human names, and I do that sometimes. But, but m- more often, like, I always have, like, the coder and the librarian and the clerk for, for things.

And you know, that, that represent usually, you know, the levels of models that I have running them, you know? Mm-hmm. L- Librarians, you know, very fancy on top of it. Um, but I'll, um W- while I'm kind of talking to them, it's really, it's really me doing the whole thing. Like, they're all mirror reflections of some aspect of me.

It, it, it's ultimately me that does the whole thing. And I, I mean, I involve other humans in this sometimes, but l- like I haven't yet worked on-- I haven't used my system in a project that involves actually multiple people. Um, partially because-

[00:59:56] John: Interesting ...

[00:59:57] Luke: with the help of these things, I don't need anybody else.

I, I mean, there's, I'll, I'll, I'll, I'll, I'll go out, you know, if, if you know me now, I'm willing to make bold, bold statements sometimes. Uh, it, it's like a John Henry thing. You, you know the, the folktale John Henry, he was like a, a digger, a coal miner, and like he battles one day against the steam shovel and, and stuff like that.

Um, I'm, I'm not that great of a coder alone. Like, just without AI tools, eh, I'm okay, but like I know, I know lots and lots of better coders th- than me. I, you know, most of the coders I work with are better than me. But with these tools, I am better than any coder that's not using AI tools. Like, put me, you know, p- put me up against, you know, Chix or Dries, and let me use these things, and don't let them use it, and I'll code faster than they can.

[01:00:57] Nic: I, I would-

[01:00:57] Luke: And more, more effectively.

[01:00:59] Nic: I think-- I don't know how to respond to that, so I think I'll hold that off. But I, I-

[01:01:05] Luke: It's quite possibly wrong, but- Nic is skeptical ... it, it, it gets like within, you know, uh- I'm skeptical ... I'm, I'm sure it's like a lot closer to them than I would otherwise be. Like, like it closes the gap remarkably.

[01:01:15] Nic: I- Uh. Yeah. Let's, I'll, I'll let that lie. Um, I'm, I'm curious about, um-

Y- y- you said that you, you've only used this by yourself. Do, do you intend to introduce this to a Drupal team at some point and test it out with other people? Or are you aware of other people that have used these tools and provided feedback?

[01:01:45] Luke: Well, here I am on a widely listened to podcast telling the world about it.

Yeah. Yeah. No, I, I mean, it's all new. This is all just a few months old. Um- I think,

[01:01:53] John: I

[01:01:53] Luke: think you'd be

[01:01:54] John: surprised at how widely listened it is, but- But anyway, continue.

[01:01:59] Luke: Yeah. No, I, I, I hope it does. I, uh... I mean, I mean, I'm- I'm humble about this i- in the, in the... Nothing I've really done is rocket science. If, if anything, if, if, if I've achieved anything, it's from having, like, maybe a, a different direction of opinions than, than other people.

Like, like, like I, I definitely historically, like throughout my career, have prioritized simplicity in a way that, that most other like techie types don't really seem to. So, so I, I, I really, you know, try to boil it, boil things down to the simplest possible thing. I, I, I do YAGNI a lot, you know? You ain't gonna need it.

Or- Yeah ... you know, don't, don't, don't put the feature in now if you're not using it. Just, you know, write it down that you might need it, and if you ever need it, then write it.

[01:02:57] John: So, uh, how do you see AMS evolving, right? So I mean, obviously the first step is getting more people to kind of, kind of use it and f- and find out where the, where the gaps are.

But like how do you see it evolving as the, the A, you know, the AI agents evolve in the future?

[01:03:16] Luke: Yeah. Well, uh, I've started actually... I, I mean, sort of the cutting edge of, of what I've been doing recently, what I did this last weekend is, is trying to come up with tooling that automates the AMS project, uh, the A- the AMS protocol.

Um, l- l- like, you could get the three agent thing, for example, called Agent Trio, by opening up three like text Claude Code windows and just naming them and interacting with them, you know, open it to the same directory. Uh, but I, I wrote a, a, um, interface over the weekend, uh, that has-- that, that lets you choose agents and will pop up almost like a little Claude Code window.

Uh, let, let me, let me go to my repo here and give you the exact name 'cause it, it's up there, you can try it. Um, repositories.

[01:04:14] Nic: We, we can, we can find that and put it in the show notes, too. I think it was called Trio.

[01:04:18] Luke: Yeah. Uh, A-AMS-Trio. My, uh... And, and that, that's, you know, it works. Um, it, it's, it, it's one of those things that it, it's like I'm in Hogwarts, right?

Like, like, I'll, I'll create a thing and then I'll be like, "Well, what does this thing that I just created do?" Uh, but, but if nothing else, A-AMS Trio will pull up a window. Um, will, will ask you to choose which personas you want to inhabit them with. Uh, I, I found bugs in it. Uh, it feels a lot like Claude Code except it doesn't handle tools quite as well, for example.

Like, like the agent will try to ask me for permission to do something and then- You know, it, it doesn't handle those messages as well as like an actual product that I hadn't written in two hours would, would work. Um, but it, it's, it's surprisingly, it, it kind of shows the proof of concept actually. It, it, it, it works as a demo.

If you, if you, you can fire it up and see it and be like, "Okay, that's kind of what I'm trying to get to." Uh, in, uh, in the Stanford talk for example, I, I came up with an interface that, uh, it, it creates a Kanban board, um, from a CSV file. And, and it, it works. Like, I was actually using it to solve a problem.

Uh, it had too much overhead, it's too slow, so you know, needs debugging, needs optimizing as such things do. But I w- I mean, m- my thing's more about user experience and UI than optimization. Like, like I can optimize, it's just not, you know, the first thing I turn to.

[01:05:59] John: Cool. Well, uh, Luke, I appreciate, uh, I appreciate your time, and thanks for joining us.

[01:06:06] Luke: Yeah. It, it's, it's been super fun. I'm thrilled to finally get a chance to be a part of this. I love what you guys do.

[01:06:14] Nic: Thank you so much. Do you have questions or feedback? You can reach out to Talking Drupal on the socials with the handle talkingdrupal or by email with [email protected]. You can connect with our hosts or the listeners on the Drupal Slack in the Talking Drupal channel.

[01:06:27] John: Do you want to be a guest on Talking Drupal or our new show, TD Cafe? Click the guest request button in the sidebar at talkingdrupal.com.

[01:06:36] Nic: You can promote your Drupal community event on Talking Drupal. You can learn more at talkingdrupal.com/tdpromo.

[01:06:42] John: Get the Talking Drupal newsletter to learn more about our guest hosts, show news, upcoming shows, and much more.

Sign up for the newsletter at talkingdrupal.com/newsletter.

[01:06:52] Nic: And thank you patrons for supporting Talking Drupal. Your support is greatly appreciated. You can learn more about becoming a patron at talkingdrupal.com and choose the become a patron button in the sidebar.

[01:07:03] John: Luke, if folks wanted to get ahold of you and talk about AMS, how best could they go about doing that?

[01:07:09] Luke: Yeah. Well, I am, uh, [email protected] if you want to email me. Uh, I'm on... LinkedIn is a great way to find me. You know, linkedin/in/lmccormick.

Uh, github/cellear. Uh, cellear, C-E-L-L-E-A-R, is my drupal.org handle. It's my Twitter handle. It's also my phone number. Uh, and-

[01:07:39] John: Interesting ...

[01:07:41] Luke: uh, it, it should be an easy way to find me. There you go.

[01:07:44] John: Should be nice. Awesome.

[01:07:45] Luke: Thanks. Uh, I'm, I, I'm a- I'm available t- to hire if, if anybody needs, you know, dev rel is one thing that I'm particularly interested in.

Uh- Sure ... product management, AI consulting integration, that kind of thing.

[01:07:59] John: Awesome. Nic, what about you?

[01:08:01] Nic: You can find me pretty much everywhere, @nicxvan, N-I-C-X-V-A-N.

[01:08:05] John: And I'm John Picozzi. You can find me personally at picozzi.com or on the socials in drupal.org @johnpicozzi, and you can find out about EPAM at epam.com.

[01:08:16] Nic: And if you've enjoyed listening, we've enjoyed talking. See you guys next week.

[01:08:20] John: Have a good one, everyone.