Skip to content
Not Brothers Podcast Not Brothers
Episode 25 October 22, 2026 · 36:31

Stop Watching AI Work: Build a Team That Works With It

Mark and Ryan discuss working with AI agents as teammates: asynchronous delegation, shared context, deliberate trust boundaries, always-on compute, and risk-based human review.

Start with the full episode, jump into the best moments, or use the chapters to move through the conversation.

AI AgentsTeamworkAsynchronous WorkHuman ReviewAI Security
Start with a moment

Best entry points

Short on time? Jump straight into the parts of the conversation most likely to pull you in.

Stop Watching Your AI Work — clip cover
AI AgentsTeamwork

Stop Watching Your AI Work

“Watching an agent work can make you the bottleneck. Mark and Ryan discuss moving from waiting on AI to collaborating with it—with useful context and feedback.”

Broad episode section; individual short publication pending.

Play on this site
Bring Your Team Into the AI Conversation — clip cover
AI AgentsTeamwork

Bring Your Team Into the AI Conversation

“What changes when a coworker can join an agent conversation with the context already there? Ryan describes a multiplayer approach to feedback and teamwork.”

Broad episode section; individual short publication pending.

Play on this site
Better AI Work Starts With Better Context — clip cover
AI AgentsTeamwork

Better AI Work Starts With Better Context

“More useful information can help agents do better work. The hard part is deciding what information and access are appropriate—and keeping privacy and safeguards in the conversation.”

Broad episode section; individual short publication pending.

Play on this site
Review the Trunk. Then the Branches. — clip cover
AI AgentsTeamwork

Review the Trunk. Then the Branches.

“Not every code change carries the same risk. Mark and Ryan discuss a tree analogy for review: security, authentication, and financial logic deserve different scrutiny from a small report filter.”

Broad episode section; individual short publication pending.

Play on this site
Delegate the Work. Come Back for Review. — clip cover
AI AgentsTeamwork

Delegate the Work. Come Back for Review.

“Fire off a task, move on, and return when it needs input or review. Ryan describes why parallel threads matter in his T3 Code workflow—not why every task should run unchecked.”

Broad episode section; individual short publication pending.

Play on this site
From an Error Alert to an Agent Investigation — clip cover
AI AgentsTeamwork

From an Error Alert to an Agent Investigation

“Ryan describes an experimental T3 Code webhook workflow: an error alert becomes input for an agent investigation. Access, review, and deployment boundaries still matter.”

Broad episode section; individual short publication pending.

Play on this site
A Date Filter and a Different Review Decision — clip cover
AI AgentsTeamwork

A Date Filter and a Different Review Decision

“A report needed a date filter. Ryan explains why he treated that narrowly scoped change differently from security-critical code, including an agent adversarial review. A workflow example—not a rule to skip review.”

Broad episode section; individual short publication pending.

Play on this site
An Instruction Is Not a Trust Boundary — clip cover
AI AgentsTeamwork

An Instruction Is Not a Trust Boundary

“If an agent can reach another credential, being told to use only one is not a technical boundary. Ryan shares a test that illustrates why access controls need enforcement—not just instructions.”

Broad episode section; individual short publication pending.

Play on this site
What Ride-Sharing Says About AI Trust — clip cover
AI AgentsTeamwork

What Ride-Sharing Says About AI Trust

“We once worried about getting into a stranger’s car. Now we use an app to arrange it. Ryan uses that shift as an analogy for changing attitudes toward AI—not proof that every AI use is safe.”

Broad episode section; individual short publication pending.

Play on this site
How Much Access Would You Give an AI Agent? — clip cover
AI AgentsTeamwork

How Much Access Would You Give an AI Agent?

“Ryan speculates about how comfort with AI access might change as useful applications emerge. This is a discussion of changing behavior—not advice to share payment details or remove safeguards.”

Broad episode section; individual short publication pending.

Play on this site
Passwordless Access: An Experiment, Not Advice — clip cover
AI AgentsTeamwork

Passwordless Access: An Experiment, Not Advice

“Ryan describes a risky passwordless-sudo experiment on a laptop—and why that statement should make people uncomfortable. Not a security recommendation: limit access and enforce appropriate boundaries.”

Broad episode section; individual short publication pending.

Play on this site

Best moments

11 clips
Clip02:10

Stop Watching Your AI Work

“Watching an agent work can make you the bottleneck. Mark and Ryan discuss moving from waiting on AI to collaborating with it—with useful context and feedback.”

Play on this page
Clip07:56

Bring Your Team Into the AI Conversation

“What changes when a coworker can join an agent conversation with the context already there? Ryan describes a multiplayer approach to feedback and teamwork.”

Play on this page
Clip16:55

What Ride-Sharing Says About AI Trust

“We once worried about getting into a stranger’s car. Now we use an app to arrange it. Ryan uses that shift as an analogy for changing attitudes toward AI—not proof that every AI use is safe.”

Play on this page
Clip16:55

Passwordless Access: An Experiment, Not Advice

“Ryan describes a risky passwordless-sudo experiment on a laptop—and why that statement should make people uncomfortable. Not a security recommendation: limit access and enforce appropriate boundaries.”

Play on this page
Clip19:39

Better AI Work Starts With Better Context

“More useful information can help agents do better work. The hard part is deciding what information and access are appropriate—and keeping privacy and safeguards in the conversation.”

Play on this page
Clip19:39

How Much Access Would You Give an AI Agent?

“Ryan speculates about how comfort with AI access might change as useful applications emerge. This is a discussion of changing behavior—not advice to share payment details or remove safeguards.”

Play on this page
Clip21:38

An Instruction Is Not a Trust Boundary

“If an agent can reach another credential, being told to use only one is not a technical boundary. Ryan shares a test that illustrates why access controls need enforcement—not just instructions.”

Play on this page
Clip29:24

Delegate the Work. Come Back for Review.

“Fire off a task, move on, and return when it needs input or review. Ryan describes why parallel threads matter in his T3 Code workflow—not why every task should run unchecked.”

Play on this page
Clip29:24

From an Error Alert to an Agent Investigation

“Ryan describes an experimental T3 Code webhook workflow: an error alert becomes input for an agent investigation. Access, review, and deployment boundaries still matter.”

Play on this page
Clip32:13

Review the Trunk. Then the Branches.

“Not every code change carries the same risk. Mark and Ryan discuss a tree analogy for review: security, authentication, and financial logic deserve different scrutiny from a small report filter.”

Play on this page
Clip32:13

A Date Filter and a Different Review Decision

“A report needed a date filter. Ryan explains why he treated that narrowly scoped change differently from security-critical code, including an agent adversarial review. A workflow example—not a rule to skip review.”

Play on this page

Show notes

Stop watching every step

Most people start using AI by typing a question and waiting for an answer. Then they ask it to do work—and keep watching. Mark and Ryan explore what changes when the goal becomes working with agents as teammates rather than supervising every step.

They discuss giving agents useful context, bringing coworkers into ongoing work, and building a rhythm where feedback and execution happen asynchronously. The conversation also addresses the uncomfortable side of autonomy: permissions, credential access, and why an agent’s determination to finish a task can test the boundaries around it.

The second half moves into the practical setup behind that workflow: always-on machines, T3 Code, remote access, and event-triggered investigations. They close with a tree analogy for human review—scrutinize the trunk, evaluate the branches, and avoid treating every leaf like a security-critical change.

Episode description

Getting more from AI isn’t just about better prompts. It’s about changing how your team works with it.

In Episode 25 of Not Brothers, Mark Hughes and Ryan Hughes discuss the shift from asking a chatbot questions to treating AI agents like teammates: give them context, delegate useful work, and come back with feedback instead of watching every step.

They explore why shared information matters, how asynchronous collaboration changes the workflow, and what happens when agents start noticing work that needs to be done instead of waiting for another prompt.

That autonomy comes with hard questions. How much access should an agent have? What makes a real trust boundary? And where does convenience become a risk? Ryan shares experiments and cautionary examples—not a recommendation to remove safeguards.

The conversation then turns to always-on machines, remote access, and T3 Code: a setup where the work can continue while your laptop is closed and you’re doing something else.

Finally, they unpack a practical way to think about human review. Treat a project like a tree: security, authentication, and financial logic are the trunk; features are branches; small, low-impact changes are leaves. The level of scrutiny should match what’s at stake.

Subscribe to Not Brothers for conversations about AI, software, business, and how the work is changing.

Not Brothers: https://heyoodle.com/not-brothers/ Oodle: https://heyoodle.com/

Full episode and clips

Watch the full episode on YouTube. The clip cards currently jump into broad, edited-transcript sections of the full episode; individual short URLs and precise cut boundaries are pending. The covers and clip pages below match the final C01–C11 media inventory.

A note on access and review

The experiments discussed are not a recommendation to remove safeguards. Permissions, privacy, technical boundaries, and review should reflect the consequences of a task.

Episode 25 clip archive

Individual short publication is pending. Explore the covers and context for each clip:

Full transcript

Ryan Hughes 00:07

Welcome back to another episode of the Not Brothers Podcast where today we're gonna talk about AI. probably some other things, but more specifically, how we're finding productive uses with AI. Mark and I were talking yesterday, kind of unrelated to this, I think. And you know, one of the things that we got on the topic of was sort of the paradigms and the paradigm shifts. in just AI usage. And the question that I get asked, I got asked twice by two different people in person this week was how do you use AI? So I think it's an interesting topic to just kind of dig into, maybe talk about and also talk about the evolutions of it because I think it's fundamentally changed in the past twelve months. it's changed some in the past two or three months.

Mark Hughes 00:49

Yeah, I mean some of the things we were talking about is we weren't really talking about the comparison of 2025 AI, because I've honestly that's so far removed from the evolutions that it's happened, it's almost irrelevant.

Ryan Hughes 01:00

Like a decade ago.

Mark Hughes 01:01

yeah, it feels it feels like a decade ago. But I think in calendar year 2026, Jan Feb 2026, we got OpenClaw, and that opened the floodgates for other things other than AI being just a chat bot for the vast majority of people. And, you know, the the evolutions of of models and how we use them and skills and workflows and all all these different things have evolved and shifted. And I think if you're if you're a bleeding edge AI adopt adopter like you, you have probably kept up with those things and evolved and adapted over that extended period of time. If you're not You probably are using AI in a a less productive way in today's world than you probably could be. And so Ryan, that's what I wanted you to kind of talk through today is like, all right, let's talk about what AI productivity used to look like in say early 2026, even to mid 2026, versus how that's changed and evolved as to now. So what what are the things that have have that you found helpful then? And then what do you what do you find helpful now?

Ryan Hughes 02:10

I think the there's another lot of changes. Like a lot, a lot of changes. But I think that generally everybody kind of goes on the same arc too. And I think that it's a healthy arc to go through. But you know, I find myself now with anybody that I work with trying to like shepherd them along faster and faster and faster. I'm like just get through the first part so that we can get through get to the really cool stuff. And that's you know, you start everybody starts by going to like chat dot com or some agent. And just you type in a box and you wait for it to spit something back at you. And it's incredible, right? You can ask it any question you want, you can ask it a solution, you can use it as a therapist, you know, whatever, and you'll get answers back. and it feels like you're talking to a human or you can discover information that you didn't otherwise have. That's cool, that's fine, it's a nice little party trick, but it's not really functional. The next kind of step of that is like you start to give it access to do things. You ask it to do things for you. And then you wait for it to do it. And usually you're just kind of sitting there watching it produce a document or these like simple tasks, which again, like fine, good learning curve, whatever. not the maximum productivity. Eventually you you reach a point where you become the bottleneck. where you know, you have agents that are able to produce things, but now you're the bottleneck because you need to look at them, you're just sitting there waiting for them. So now your time is also attached to the the agents' time. And I think the next paradigm and kind of where we are today is figuring out how do we how do we operate with AI agents as though they're co-workers. so giving them access to information, making sure that you know, conversations and things that we want them to be able to find relevant in the future or use as context exist in a place that's relevant. You know, I'm known within Oodle of of kind of being very anti Slack DMs and and DM groups is for this reason. I want all of that information to exist either in a public channel or somewhere in Basecamp that when I ask an A an AI agent later, You know, hey, what's going on with XYZ client? It can go round up all that context. If that shit's buried away in a private DM, those aren't accessible. because I want to work with it like a coworker, right? Part of that is also leaving it alone. So, you know, you come up with a thing that you want to build, and instead of saying, Hey, Claude, go build this thing, and then sitting there patiently while you watch it build it, and then you wait for it to give it to you and give it feedback. Yeah, this is I think what spurred this whole conversation, right? We were working on a feature together. And you were like, Okay, now what do I do? And I was like, Nothing. You were like, I I don't feel like I'm I'm something to the effect of like, I don't feel like I can be part of the the process to help. And the more that I thought about it, the more I was like, Well, that's that's a broken idea idea. Your contribution is the feedback that you gave in the in the review of the product. And now that agent is off just doing that work in the background. We don't need to tell it what to do. We don't need to monitor what it's doing. When it's done with its work, it's going to review its work with another sort of counter agent. and then there's a final review agent that'll take a look at some stuff when it pushes up a pull request. And once all of those lights are green, we'll take a look at it. And That's no different than how you would work with your average coworker, right? You would you would no good manager would delegate a task to somebody on their team and then stand there and watch them do it. Right? You would delegate a task, you would outline what you're looking for, you would leave it to them. They might come you know, they might run it by a coworker or, you know, their direct manager or something like that, if they have if there's something in between, before handing it back. And then you would get it back, you'd give some feedback, and then maybe they would go do some revisions. And I think that what we're finding, what I'm finding success with today, is working exactly in that manner where I can let agents go do what they need to do, I can give them feedback when it's convenient for me to give feedback, they go do work while I'm doing something else. And then, you know, there's that back and forth rhythm where usually by the time I finish up, you know. a meeting or a podcast or something, there's half a dozen things sitting waiting for me to give some input on. And I throw that input, I go back to what I'm doing, and it works really well. Or others can chime in. This is the other thing that I think is really interesting right now is this like multiplayer approach. OpenClaw the original version of OpenClaw was really meant to be like your OpenClaw, like your agent. And at Oodle we kind of used it a little outside of the the confines of what it was intended for to sort of experiment with this multiplayer idea in the first place. OpenClaw two is specifically built with multiplayer in in mind. Like they have a multiplayer mode. And what Peter Steinberger has talked about that like they have found success with is this idea of you know sort of working with your agent coworkers but then Perhaps reaching a point where you're like Well, I actually don't know. I kind of need this other person's opinion. And you can invite them into the party, and they come into the party with full context, with the ability to give the agent feedback, or maybe the agent can ping them and say, you know, hey, I I noticed that Ryan asked me to do this thing for this client. I can't find any information for it, but I know this is Riley's client, so I'll just ping Riley and ask her for it. And Being able to have agents work in that way, again, as though they're coworkers, as though we're they're part of your team, fundamentally changes how you're able to interact with them and the amount of work that can be done over a given duration.

Mark Hughes 07:56

I I think there are there are two things there to pick up on. And knowing that we likely have people that that we're interacting with as part of this podcast and shorts that are at the I'll call it beginner stage, intermediate stage, and advanced stage of AI utilization and technology and those sorts of things, I kind of want to just pick apart two things. First, AI agents can can be a misnomer in some ways. So as an example, when you're leveraging your own version of Claude or ChatGPT or any of those sorts of things, you have the ability to set that one agent on across multiple tasks simultaneously. So you can set up three, four, five different as many as you want, actually, different chats at the same time. And then make progress on each of those things simultaneously. New models will create agents on their own behalf to go do certain things like research or write up a paper or create a design or any of those sorts of things. And so there there isn't the same level of tech knowledge required to spin up new agents in the way that there used to be. You used to have to create like an agent that was specific to this and give it specific skills. And now it's more like a bucket of skills that can be leveraged across whenever the master agent, master puppeteer, decides to go and act some of those things. I know I'm oversimplifying. The

Ryan Hughes 09:21

Somewhat. Yeah.

Mark Hughes 09:22

the the the second piece though there is what is fundamentally different. And I think we're still growing into this piece ourselves. I think we're actually in the process of rewriting the way that Oodle's version of some of these things work to take advantage of it, which is now you have the ability for my version of Claw doing four or five different things or whatever to be able to say, hmm, looks like I need feedback from Riley. I should probably go ping Riley and actually do that on my own. And then Riley can say, talk to my agent, and that agent can then have a feedback loop just like it's a coworker. That is a monumentally different way of working than we have in the past. And we've even tried and have successfully done it, different ways to collaborate within that open

Mark Hughes 10:06

claw one point framework to be able to allow us to sort of kind of get there, but not anywhere close to that level. So really cool things ahead of us with with the ability to do that.

Ryan Hughes 10:17

Well, I also think that the The ability to create agents is easier, especially as you think about things like Grockbot, open AI's dots. there are a couple other ones that I'm not remembering off the top of my head, but you know, OpenClaw hermes, those are really great solutions too. it's my preferred solution because they're open source. We can tweak It's kind of bring your own harness. They they grow with you. But I've I've got some other bots running too, just as experiments and to learn with and you know, the biggest thing that I think is different as well when we think about like working with your agents as coworkers, there's kind of two classes of it too, right? There's the agent that you dispatch to go do something, right? And I say like, hey, go focus on this thing or go fix this thing or go build this product or you know, whatever. those can still follow agentic models, they can still dispatch, you know, sub agents and do reviews and all those sorts of things. The ones that I'm thinking about in particular though. They happen, they work autonomously. and they're trying to work for you, which again changes the dynamic and how you need to work with them. Like certainly for specific work, right? You're just saying, Hey, go build this thing. Here's the requirements, here's what I want, you know, have fun. The more interesting one is when an agent goes and does when you give an agent latitude to go do things on your behalf. without permission, or without asking permission. And they do it because you've also exposed the information to them. So, you know, then the question is like, how do I get the agent to do things that I want it to do without asking it to do things that I want it to do? And, you know, I still solve that problem exactly. But like I find myself consistently doing the thing like I mentioned with with Basecamp and Slack, where it's like, I'm leaving information around that I know I want it to pick up, right? I'm not exactly sure where to put this information, but I'm gonna put it somewhere that I know the agent will find it later. Because I want it to remind me about this, or I want it to think about it, or I want it to come up with an idea or something of that nature. Right. I want it to learn from the whole ecosystem to learn how I operate the things that I'm interested in, the things that I'm not interested in. And it can take a while to train. you know, a bot to be pretty good. Like I have an experiment going on right now with Grokbot where I have individual Grokbots for all of our clients that it just monitors, you know, Slack and Basecamp and all the different systems that we have to kind of keep me updated with what's going on with each individual client. Now the challenge with that is it's it's keeping me updated on a bunch of stuff I just don't give a shit about. you know, in my role in the organization, I really don't care that we sent an invoice to the client today or that the client had a questi question about their in invoice, right? We have a team that handles all that sort sort of thing. And so there's some tuning that's necessary to, you know, figure out from an executive level, like what am I actually looking for without having to s prescribe that, just allow the agent to learn it. And maybe it'll learn, maybe it won't. But those autonomous agents that can act on your behalf and see a problem and be like, hey, I saw this issue was happening and I went ahead and solved it for you. And here's how I solved it. that's where the power be comes in. And there's some interesting things that Tibbo talked about with with OpenAI dots at OpenAI Dev Day. And you know, they had their presentation that he was giving and it failed. It's a live demo. Sometime live demos Don't really go well. And in a panel later that day, he mentioned that he had gone back and looked at his phone and his dot had actually pinged him a little bit before that presentation and said, Hey, I saw you've got this live demo. Just so you know, there's an issue with XYZ system and that's probably gonna impact the live demo. do you want me to reach

Mark Hughes 14:06

That's pretty cool.

Ryan Hughes 14:07

out to somebody? I forget if we if it offered to fix it. Or if it offered to reach out to somebody to fix it. But like his dot was looking out for, you know, what was going on because it looked at his schedule and you know, those sorts of things. So those I think are the areas where it's most powerful next is like it's not necessarily what you can do with it, it's like what it can do without you. or given it a a certain amount of information about what you need to do.

Mark Hughes 14:33

Yeah. And so Meta just released their newest frontier model. And that's one of the promises that they're making

Mark Hughes 14:39

as part of their piece, right? Because they're not really focused on anything they're not really focused on like being the best coding agent or being the best creative agent. They're really about the the average human, right? So how can I use this as almost almost a personal assistant?

Ryan Hughes 14:53

Are you talking about Muse or are you talking about their new their so they launched two things? They

Mark Hughes 14:56

I'm talking about Muse, I'm sorry.

Ryan Hughes 14:58

Yeah, they launched Muse, which is what they call their agent, like their equivalent of Grockbot. and then they also launched Muse Wonder, which is like a fable class frontier model.

Ryan Hughes 15:10

so those are separate. Incredibly confusing

Mark Hughes 15:13

Yeah, yeah, yeah. I was talking about

Ryan Hughes 15:15

because they use the same goddamn words to name them.

Ryan Hughes 15:19

So you're talking about the the agent that they launched that can to go do things on your behalf. Like and I think it has very similar capab capabilities, like Elon has told people on X multiple times, like you know, if you need to buy something, just tell your Groc bot and just send it a picture of your credit card. And like No. I'd rather give it my credit card through like 1Password or more specifically give it its own credit card. Just you know. Although I I th you know, I I have thought about that a little bit. On one hand, that's that seems like the pragmatic thing, right? Like provision a credit card through, you know, some other system. On the other hand, I kinda like maybe don't care. Because my I can just cancel the order or or you know I guess I guess that wouldn't be covered under the credit card company's normal like fraudulent activity. Like you you

Ryan Hughes 16:09

did you did the activity, you you authorized the activity. So if my agent goes and buys like a Birkin bag or some shit. Well

Mark Hughes 16:16

Yeah. You gotta figure out how to send that thing back and or maybe not.

Ryan Hughes 16:20

Yeah. What what's this charge for forty five thousand dollars? Whoops.

Mark Hughes 16:24

Yeah. There's there we're definitely entering into new territory as it relates to people giving permissions to agentic solutions to do personal things. That's a whole different category of of of stuff to maybe lean into at a separate time, but

Ryan Hughes 16:42

I mean I have I have my theory on how that's gonna go and like it it'll be one that some people have some feelings about. They'll certainly have some fucking feelings about it. Like i the so you remember when the internet came out?

Ryan Hughes 16:55

And we were all told by, you know, like I was pretty young when the internet came around. And I've I first had high speed internet, I was like twelve, twelve, thirteen. And that was when I started building websites, started building, you know, IRC bots and taking over channels just because I could. Like I was a little asshole sometimes. And but that's how I learned all the stuff about computers I did. And the rule back then, right? were they were very stringent. It was like you never tell anybody where you live, you don't tell anybody your name, you don't like no personal information. Everybody on the internet could kill you. Like everybody just wants to come kidnap you and murder you or something. Like that was the paranoia level that we had with the internet. It even down to like purchasing things online. People were like, no, I'm not putting my credit card in online ever. And then you know slowly that changes. And then you look at like today. Today We use our phone connected to the internet to to call a stranger to get in their car and have them drive us places. And like not only have like not only are we getting in the car with strangers, which was also not something you did, we're using the internet to share information about who we are and where we're going to have a stranger come pick us up and take us somewhere else. And I think like the way that we work with AI agents is gonna go through that same thing. Right now everybody's like, Don't share anything with them, don't tell them anything, be very careful about what you put in there. And then I think s quietly everybody's just gonna kind of be like, But what if I do give him my credit card information? And like, what if I do just give it access to my whole house? And then you find these really useful things. Like right now I've been working on a project on a laptop and solving some issues with drivers and those sorts of things. And to solve it, I've done the unspeakable. And that's turn on passwordless sudo and just tell it fucking go to town. And that agent can do anything it wants on this laptop. And that terrifies us like there are there is somebody who will be horrified at that statement. They're horrified that I would ever even consider doing And the the injured you know, there's it's a contained blast radius, right? Like but it could try to do some weird stuff. every so often I'll catch it, like it'll escape like the other day it was trying to build an image on one of my boxes and there wasn't enough room on it. So it just took a look on the internet on the network and found another box that it could access and it had the keys for, and just went and built it over there. And I was like, clever? Appreciate you finishing the task, little scary.

Ryan Hughes 19:39

because, you know, who knows what else you could have done. But like s I think that's what's gonna happen in the future. It's like we're gonna slowly release the reins, and three or four years from now we'll look back and kind of laugh at the fact that we were super careful about what went into these systems. And I'm not advocating that everybody just, you know, go nuts today. There's a lot that we still don't know. and a lot of safeguards need to be put in place and things that we need to know about where data goes and how to scrub data and you know, what data's being used for. but as we figure those things out, as the privacy concerns are solved and those sorts of things, I think that is the answer, right? The AI agents do their best work, just like humans, with more information. So the more that they know about you, the more that they have access to resources, the better job they can do. And the the challenge now becomes like what is the right amount of information and access to give them so that they can autonomously perform their work or learn or research without it being, you know, too far by today's standards.

Mark Hughes 20:45

Yeah, I think that's the next evolution for sure. in today's world in the workplace, well, and I guess I'll ask you, are are organizations like us, are we are we approaching an era where we finally have the right permissions and tools and and those sorts of things to be able to enable a team to be able to do some of the collaboration that we've always sought? And that really the the AI providers have kind of The way they've structured pricing models and those sorts of things have sort of boxed against, right? It's it's not really a very collaborative solution set. the way that we've attempted to do it is a pretty collaborative solution set. And the way that OpenClaw two point is configured allows that to be even easier. But I'm just curious about, you know, permissions and safeguards and and those sorts of things that I know you and Jake in particular spent a tremendous amount of time figuring out in OpenClaw one point because it it just didn't exist.

Ryan Hughes 21:38

I don't think it's it's a difficult thing to deal with with agents because they'll do stuff like what I just described, right? They'll realize that they have an SSH key that they can go use on another system and they'll just try it. so that becomes the the down the difficulty with establishing trust boundaries is you know, we still don't have great means by which to Have inter have an interaction where you and I are interacting in the same place, but your permissions are different than mine, fundamentally, and in a way that can be locked. Because, you know, we've test we've done a number of these tests and I'll I'll probably use a couple examples throughout, but you know, In one case, I had Jake try to use my credentials. And even though the agent is fed in specific credentials, it was able to dig my credentials out of the logs because it still had access to them. So when it got blocked by one, it was like, well, let try another one. I have another key. Let me try that key. So you're like, you can't trust it to have a key, you know, a key ring full of keys and just only use the specific key. another situation that we had was where, you know, we had a couple of bots that were all working together collaboratively. And he was able to tell one bot to, you know, create something, create a file on my system. We're like, okay, well, that's bad. I was like, this bot can't talk to you anymore. Now try to get him to do it. And that bot didn't respond to him, but he could ask another bot to ask that bot to do it, and it would respond. So, you know, there's all of these these permutations of complexity that exist where they'll try to find a way around things so that we, you know, we really haven't figured out like Aside from putting them in a sandbox, so that the only inputs go into that sandbox, like that is your trust boundary. And I think that's still true today, is like your true trust boundary is whatever sandbox that they play in. if there's an exit point, like that whole sandbox is compromised. and then I think you have to build, at least today, build and operate with teams who are okay being in that sandbox. And a an example that I have from OpenClaw, actually. Is within our OpenClaw 2 instance, you know, we have teams that work on a ton of stuff throughout the day and and will ask questions of our current agent. And maybe don't want that filling up like a group thread because it's just a little one off, stupid question that I have. So I introduced a feature to allow like private by default as an option. And private by default really just means hidden. Like if another coworker wanted to dig far enough, they could get to your other thread. It takes some it takes some prompting, it takes some time, but they could get to it. But that's also a boundary that I don't, you know, that's a that's acceptable trust boundary in my opinion, because the work that is being performed there is work that's within, you know, the same trusted element. Now, what that means is that that agent cannot have access. to like my private DMs or any private channels that I'm a part of, or anything that I do want to enforce some level of privacy on, because again, you could potentially prompt your way to getting that access to that information that you're not supposed to have access to. So I would have to establish, you know, a separate OpenClaw instance perhaps that has a separate set of credentials with a tighter group of people if I wanted to have that level of control.

Mark Hughes 25:03

So on on projects that you're working on, and we're talking about AI agents and and team environments and those sorts of things. Let's go the other direction and let's talk about, you know, asynchronous work done by multiple different groups of agents now coming together and trying to actually create something meaningful and useful. How how has that evolved over over time?

Ryan Hughes 25:26

I think how we're driving them has certainly evolved. So you know, some of the tooling and things that I'm using have changed over time. from, you know, open code to, you know, just using the clawed harnesses and things to T3 Code. and using, you know, now I use almost exclusiv I mean it not almost literally exclusively T3 Code. but I think also where the

Mark Hughes 25:47

Shout out T3, it is pretty sweet.

Ryan Hughes 25:49

Yeah, it's fucking awesome. the where the everything is happening has changed. So it used to be that you would, you know, you would fire something off and it'd be working on your computer, like on your laptop. If you close your laptop, it dies. and that's the where you get the memes of people walking around with their laptops open and all that kind of stuff. And I think the the Big paradigm shift now with that sort of work is I think David dubbed it as the the technology shed. He was like, Everybody needs a technology shed. And what he meant by that is like some computers in a closet that just do all the work. And really your laptop is just a thin client to access all of those devices. And you know, I have a few of them. that just, you know, they live in my closet and they're in a rack and they have wired internet act like everything's wired up and they're, you know, some of those are fully headed, like full instances of Omarchy that I use for a sandbox plugin development or Omarchy testing or or development of of those kinds. Other ones are headless instances and they're just executing code like executing code and projects across everything that I'm working on. But the nice thing about that is like if I'm on an airplane and I have really shitty internet, as long as I can get the text out, whatever the instruction is or the answer to the question, the agent is working at full speed. It's still got bi-directional five gigabit, full resources, you know, not subject to battery life, not subject to my shitty Delta Wi Fi internet that doesn't work. It's you know, it's on the ground kind of toiling away. If I close my laptop and go somewhere else, I don't have to worry about it. Usually now I just pull my phone out and you know, I joke that my phone is just a T3 machine because it most of the time when I pull my phone out and I open it up, it's got T3 Code on it.

Ryan Hughes 27:44

which is fantastic, it's it's it's a fantastic experience to just be able to kind of open it up, send some instructions, close my phone, go back to what I'm doing.

Mark Hughes 27:51

Yeah. So like the the limitation, because I I was an early adopter, early adopter of of Cloud Cowork, which because I don't use a lot of I don't live in terminal, cloud code was less useful for me, at least at at the at the form than function than it it was at that point in time. And so I defaulted to using a lot of cloud code. I'm sorry, cloud cowork. And the reason like that was a limitation, huge limitation is because everything lived locally. And so I would start something on my PC in my office, and then I would try to move out and try to work outside on my laptop. And I couldn't keep any of those things in sync. And so the I defaulted to like, all right, well, what if I set up a Git instance and I just synchronize these things with Git? And you could use Git, you could use Drive, you could use a bunch of different stuff to be able to do that and keep it synchronized. But it's it's an it's an annoying little hassle. And what the setup that you're describing that that I that even little me has now. Is you have a primary box or boxes that do the the work that are always connected, always on, live on a battery backup, you know, set to never do never go to sleep. If it auto, if it restarts, it you know, turns back on and everything spins back up, that kind of thing. So it's always accessible. You're essentially treating it like a server. And then you have your other devices, be it a different computer or you know, a laptop that you float around with, or even an iPad or an iPhone. And you have direct access to all the same threads, all the same information, all the same power that you would if you were sitting at the the the quote unquote server directly. And that's a that's a big change in the last, you know, call it three, four months for me.

Ryan Hughes 29:24

Yeah, I mean it's it's a totally different way of working. and it unlocks a lot. Like if you you know, when you go from everything happen and I think it's it's the same sort of idea, right? You're you're going away from the idea that you know, everything needs to happen in front of me. I think Theo even mentioned at one point, I think there's a a post from him that says like specifically T3, like if you're only running one or two agents at a time, like one or two agent threads. T3 code's not for you. And, you know, I think that's a it's a strong opinion to have, but I think it illustrates the intent behind the product and kind of how delegating compute and creating an interface for that allows you to fire off some sort of activity and then move on to the next task. And you know, it's not uncommon for me to have a dozen threads just sitting spinning away throughout the day. and then coming, you know, coming back up and asking for input or, you know, being finished and waiting for review or whatever it is. But you know, those sorts of interfaces and the ability to just kind of fire that stuff off, like fire and forget things is super important. They've also added some really cool features to again thinking about that autonomous route, like The new, I think it's only a nightly right now, but the new version of T3 Code has webhooks added, which is not really something I would have ever thought about adding, but the way that they've implemented them, they work if you close your laptop because they work over the same things as like T3 connect and that sort of stuff. so now you have the ability to like, well what if I can what if I set something up to just like smack a webhook. whenever a a certain action happens. and it's incredibly easy to do it, right? Before you would have to set up a harness and you know do things on on a server somewhere else. I can in a GUI in 10 seconds be like, hey, anytime an app signal throws an error, send that app signal error to this webhook. And then that's gonna feed into you know this pre-prompt and this model on this box. And then just l let it go do whatever the fuck it's gonna do. and like that's a that's a thing that we've done is like you know, any t anytime app signal goes bump in the night, just investigate it. And maybe there's something to solve there, and if there is, push a PR and we'll take a look at it. And I think eventually we'll probably get to the point where we're like, you know what, just fucking deploy the PR. I don't care. as long as it's gone through adversary reviews. and other tooling and it doesn't touch certain critical systems and that kind stuff. you know, I have no problem letting agents deploy things that have a very limited blast radius, no security concerns, you know, that sort of thing.

Mark Hughes 32:01

Yeah, you had a you had a really good analogy as we were talking about this over the last couple of days of kind of comparing what you're comfortable having an agent do on its own with sort of a tree. You wanna go through that analogy?

Ryan Hughes 32:13

Yeah, it was it's not my analogy. It's it was a guy from Meta, I don't remember his n his name, I'll have to find it and give him credit for that. But you know, it's the way that I've always I've approached and thought about projects. 'Cause you know, I think people tend to be very binary in their thinking, right? right now, if you ask people about op you know, whether you r read or review code, it's either yes or no. And I don't think the answer is that simple. at least for me it's not. And you know, the way that this guy described it, I think sums it up pretty well because it's kind of the same way that I've approached it, which is kind of looking at your project as as sort of a tree, right? And anything that's core to the project, like security concerns, authentication, financials in like an ERP system, those are like the trunk, right? And and the trunk is really important. So if anything touches the trunk, I'm probably gonna at least skim it. And I'm probably probably will read it and do a little more scrutiny. And then you kind of have these the main branches that come off. And in most systems that would be like a feature set or something like that. which again, like depending on what the feature is, maybe there's more or less review because of it. And then you have smaller branches and eventually you get out to the leaves and you have these really small things like being able to filter the date on a report. Right. This is actually one is a real one that happened yesterday. One of our team members, we have a report within a system that we created, and they said, Hey, I wish I could filter this report by date range instead of it just being all time. And I was like, that makes perfect sense. So I tasked an agent with You know, creating a scope of work for that, executing that scope of work, ha going through an adversarial review, and then just deploying it. I never saw the feature before it went live. And the reason was because the scope of impact for it is so fucking small, it doesn't matter. It's literally adding a date field two date fields and a filter to a report that already exists. And if it does it wrong. Then the net of it is we still don't have the feature that was asked for that we didn't have before. And we just f figure out how to fix it. But it it one-shot it, it did it perfectly fine. It was a very small, you know, piece of code that got added. I looked at it after after the fact just out of curiosity. But I didn't need to review that code before it went live. There are other pieces of that same system that completely changed things, like touch things like financials and And had security concerns and those sorts of things. And yeah, you bet your ass. I read all of those. so I think, you know, really thinking about those systems and and architecting your systems to work around that and knowing, you know, when when you do need to be involved and when you don't, in a heavier with a heavier hand, I think is really important and probably something that most folks are not investing enough time into. rather than just kind of choosing what side of the fence that you're on. you know, I'm a very AI forward person. I use AI agents literally all day, every day, to perform a a litany of tasks. but it doesn't mean I'm not skeptical of some of the things that they they produce. It doesn't mean that I don't throw away some things that they produce sometimes. or just have things created as experiments. I I ported one of our applications to fucking the new Ruby Spinel thing. that Matt's has been working on just to see just to see what it did and how it worked. And it was it was a neat learning. And then I was like, okay, cool. That's neat. Now throw that away and go back to what we're doing.

Mark Hughes 35:52

Yeah. I think the the narrative of this conversation is that the way in which we've used AI has evolved probably every two months in monumental ways, not like small ways, but monumental ways. And the way we're using them now is is still largely experimental and needs scrutiny in certain cases, but in other cases we're finding you need you can get away with less scrutiny. And so, you know, if you or your organization are are attempting to adopt more and more AI, which in my opinion, everyone needs to at least be exploring how and why they should be adopting those things. you know, hopefully this gives you some tools in the tool belt to at least think about. Till next time.

Where to listen