Your AI Assistant Should Act, Not Just Report Vibing with AI Code, Episode 14: Brendan & Baird Brendan: so Grok Bot came out recently, and it's been really making me rethink, uh, my Claude Command Center you know, just more or less like how I interact with it and what else could be doing with that to, you know, just, just make it more… Like, useful is not the right word, but it's kinda like a good point for it. A- and, you know, I, I've had this feeling now for a while with it, and I've been using it for around six months or something, that there's something missing. And it- while it, it has been helpful and it definitely has made me more productive, I'm still not feeling as productive as I, as I, I should be. Or another way to put it would be just, like, other people that I see or talk to are using AI in ways that it feels like, at least on the surface, they've broken through. And, Baird: Sure Brendan: w- with Grok Bot coming out, um, I recently had a discussion on my other podcast, the Rails Business Podcast, about, uh, with Jesse Hanley, who's from Bento, and he is using Grok Bot, and he is being very productive with it. And so it's-- That's kind of where it all kicked off, and then, you know, I've explored Grok Bot from a high level a little bit. don't really wanna make the transition, and that's the other part. It's like it feels like… I, I should say I don't wanna like do a complete shift away from what I've already built. But I do want to, you know, make what I built more useful. So I figure we can maybe talk about that at least a little bit here and just explore it and, um… Yeah, I don't know. Any thoughts before I kind of dive into what I've done so far? 'Cause I haven't done a lot, but I have started sort of reworking things. Baird: Yeah. I think from my perspective, I'm curious, like, what isn't feeling productive from your perspective? 'Cause, like, it's something that I've seen on my end too. It's like, if you look at all the work I do some days, it looks like I'm being super productive, but it isn't always the right type of productive. A- and I think that's kind of a, a key thing to separate is like, are you doing a ton of tasks that are high value or are you doing a ton of tasks that are, you know, not so much? And, and like how do you prioritize that? So I'm just kinda curious, like, where your head's at on the right type of productivity. Brendan: That's a really good point, and it's, yeah, it's a valid question for sure 'cause like It, it's a bit of a mix, I guess, is part of the answer. So I, I feel like in some ways the code review stuff that I've been using it for worked out be- let's put it this way, it's been better than before, but even that, like, even that specific thing, there's still some, some edges, I guess I'll call it. But I'm feeling pretty good there because I already have the whole system, like, going out, it kind of does, like, a pre-review, it brings it to me, and then I, I guess that's kind of where it breaks down a little bit, is that already done an initial review, but the way the system is working today, it feels like I'm doing some rework where every time… It's almost back to that interface thing, where, like, because I'm interfacing through th- a chat agent, I'm not… I don't have an understanding of, okay, which ones were, were there? Which ones are ready? Like, what was the review? So it's like I'm re-asking questions that I don't necessarily need to. So in that c- that case, and I don't want to get away from your question, but just to, like, illustrate that case, you know it feels like maybe there's, like, a UI layer or something that would, would make me more productive with the Baird: Sure. Brendan: getting, Baird: Yep Brendan: So that's one part. But I think that's, like, the part that's going the best. The part that's struggling and where I'm, I'm trying to figure out, like- Am I either not being productive or w- how can I be more productive, I guess, is where, what I think, but maybe it is a little bit of I'm not even thinking about the framing quite right. But to just paint the picture, you know, I've had a morning briefing thing that, that was, like, one of the first things I set up because, kind of because it's so easy to do. You can just, Baird: Yep. Brendan: there and you can hook your MCPs up to your email and your Slack and your GitHub and whatever, and then you can tell it to grab a morning brief and then, you know, we've talked about this before. It becomes a bunch of noise and you stop paying attention to it or you, it, whatever. A- and through this whole process, by the way, like, my Claude command center thing told me that we've tried to rework morning brief alone six times already, and so Baird: Sounds right, yeah. Brendan: there's something else going on. Baird: Yep Brendan: but I think what I… Part of it is, like, when I do, and I've, I've almost completely stopped paying attention to the morning brief, and so much so that, like, when my Mac Mini stopped running it for the last week, I didn't notice because I Baird: Yeah. Brendan: wasn't Baird: always a good test. It, it's like i- if it fails, do you notice? Brendan: Yeah. But when I was getting it, you know, more often than not, it, it felt like there were gaps, too. I guess that's the other part of it, is even though I've, it already has access to all these different things, it would be telling me that I didn't do something or, like, this is your top priority and it's like, no, no, no, I did that. it's in Slack or at least it's in some place that you could, should have or could have found. And so that's the part where primarily it felt like I keep having to almost fight the system to get it to know what I'm working on y- you know, unless I specifically sit there and be like, "I did this," and log it. Which honestly was something I did for couple days just to see, like, does that make it even any better? No, it Baird: Yep. Brendan: But, Baird: Sounds right Brendan: the, I guess the other thing to just add is, like, it… So and I guess this has made me realize, or at least, you know, because of your question, it's not that I feel like I'm not being productive. It just feels like I'm still missing things, and the system, while it does help in some ways, is in s- also in some ways it's, like, another barrier that, uh, is making it slightly worse possibly in Baird: one more thing you've got to manage. A- and so it shifts the responsibility from, like, four channels to one channel. But when you've got that, like, gap in context, it then makes it, like, almost more frustrating sometimes. 'Cause it, you know, I found it with the agents stuff I've done, it's like when it works well, it's like, oh, this feels really good. But when it goes wrong, it actually feels almost worse than just managing the other stuff independently. You know, 'cause it's like then you at least have your context which pulls it all together Brendan: Yeah, that, that's a good way to put it. Yep. Uh, you know, and two, it's like it is that context problem, right? Where, where even with a coding agent, if the, you're writing code it comes across something that you know what it is because you wrote it before AI was even writing code, and there was a reason you did it this way, but you, now you've told AI to build whatever it might be, and it comes across that line of code that on the surface looks like, we gotta change this," or, "Oh, that's something important maybe." And this is even the better example, is, oh, that's something important, so instead of it being a one-line change that Brendan knows because he's got the context, becomes AI trying to j- achieve that goal that I gave it by writing an entire new file that, like, goes around that code Baird: Yep, for sure. A- and I've even noticed that as like, I mean, it's a great example of like if a human wrote the code and then the AI sees it, but I've also noticed it if like either a different model wrote the code or even like a different session that never, you know, put decisions either in comments or in the agent's MD, it can do the same thing where it's like… And then it misattributes who actually like made the suggestion. Like, I always love when Claude's telling me I made a decision when it was like totally its direction and I'm like, Brendan: Yeah Baird: n- no, that was, that was actually you," but, but no, I, I totally get what you're saying is like, unsurprisingly, it is a context issue. And, you know, I'd be curious, you know, with the agent stuff… Brendan: of it's a context, context issue for sure. Yeah, I think there Baird: No, it-- Brendan: talk about the context stuff still Baird: Yeah, like I, I agree. Like the whole issue you're talking about isn't solely a context issue. Like the gaps in the usability stuff is like the not knowing if something's done. 'Cause like that was one of the things I very purposefully and, you know, why I was playing with the agent I was, been building is like, it's harder to make those rules in a Claude Code, you know, routine, versus kind of having your own agent framework where you can truly bake that stuff in. But I've noticed it too, where it's like I even have, you know, a index set up so my agent can actually search for existing things that I've done and actually like compare. And it's still, you know, I use Linear to try to track everything, trying to like figure out like what's the right UI layer to help with this. And I still find it's like putting duplicates in, where it's like because there's like a missing piece of context between a granola note and a Slack, or a granola note and, you know, a email thread, or even just like stuff I've put in. A- and so there's still, even when you start making it like more custom, there's still these gaps that just make it really annoying because then it is just like… I mean, if I look in my Linear board right now, and I've probably got like 75 different things sitting in there, Brendan: Yeah Baird: you can group them down to like probably 10 different things, you know? And so there's just like a lot of weird noisiness. A- and so I totally get where you're coming from. A- and I think there is kind of that open issue there. But I'm also curious, you know, what about Grok Bot kind of sparked that it could be better, like that that's the path to explore as the next potential iteration? Brendan: Yeah, I think the biggest thing there, and this is something I totally could do with anything, right? It, it doesn't have… I mean, OpenClaw's been doing this sort of thing. Hermes has been doing this sort of thing. You could make Claude Code do it, and that's what I'm exploring now. But what it is, is it- it's not even the fact that it's, like, all different chats essentially. It's more, I think, to, to do with the fact that, like, I have never felt like my system was doing enough for me. Baird: Sure. Yep. Brendan: a lot of things. It's creating Baird: Yep Brendan: of content for me to look at, but it's not, Baird: Yep. Brendan: it's Baird: It's not acting on your behalf. Yeah Brendan: write. And that's what I, I, I basically, you know, went to it, it was just yesterday, and I'm like, "Okay." And I th- actually, I think I'd already brought up just, "Hey, we should, like, rethink things." But w- instead of just saying, "Let's, hey, let's do Grok Bot instead," the way that I went about it was I said, "You know, it's been six months. Like, this has been an improvement, but also there's been some problems. So, know, you've got all these MD files now. Get, get the context you can get out of this to try and help me understand, like, at least what your thoughts are and where you think it's breaking down." And then I also just layered in, like, you know, "I know people that are using things like Grok Bot, and they're being successful at it, so you know, just consider that as something we could either move towards or, like, in our system, I could just, like, you know, completely cut off where I'm at here and go onto that completely." Baird: be prepared for the blackmail, 'cause then they do the, the research, then the club will, will resort to blackmail if you say you're gonna turn it off. Brendan: I haven't seen that, but it did tell me Baird: Okay Brendan: the lines of, you're asking Claude to move away from Claude?" I was like, "No, I said, "If Baird: make yourself valuable. Come on Brendan: Yeah, right. Yeah. But yeah, I, I think that's the biggest piece that is almost completely missing, is just the, like, it closing loops for me. You know? It's like, if that co- I- and this, this is where, to me, it's always felt like, okay, the context is there, either I'm not prompting it well or I just need to be… make sure that its understanding is, yes, go look until you find it, 'cause I know it's in Slack, for instance. You just need to go find it. So, you know, a good example would be somebody sends me a message in Slack, and, or even an open poll request or something like that. You know, someone requests me to do something, and that artifact lives one of these things that Claude already has access to, then I have acted on it, and the completion of that is also in one of those places. And there's been several times, again, where, like, my morning brief would say, "Oh, but that's not done," even though it was done and it was in one of those places where it could, i- in my opinion, it should have found it. Um, and so that's one of the things we're starting with. And, you know, we're not gonna just go straight to, how do I build, rebuild Grok Bot in, in, uh, Claude Code? It's, it's like, here are the gaps and the things that have been a problem with the, the current setup. do we be more proactive? So, like… I also heard, um, a podcast yesterday, uh, My F- My First Million, with, uh, Wade Foster, I believe is his name, the CEO of Zapier. Uh, and by the way, I know a lot of people call it Zapier. He, he did a g- he had at the very beginning a, a good way for you to remember, that it's Zapier makes you happier. That's how I remember. And honestly, I've al- Baird: Fair enough Brendan: it Zapier. We're getting off topic, but- The, the interesting thing, the reason I bring it up is 'cause he was talking about having a lot of these same types of things that I've been doing with my command center already, but he wasn't talking about it, i- you know, it was all positive. you know, it was more like when having, when I had that conversation with Jesse about like, "Oh, it's making me more productive." You know, his morning briefs, 'cause he has a morning brief, he's like, "It's great." then the other reason I'm bringing it up is because this is the thing we are currently working on, and it- what he said he does he has… The more valuable thing is he has a, an evening brief or, like, an afternoon Baird: Okay. Brendan: and a Baird: Sure. Yep Brendan: that alone, like, does a good job of kinda summing up, like, all the things that he's done during the day, but also like, you know, it's kinda like his to-do list going into the next day already, Baird: Sure. Yeah Brendan: and just closing the loop on things. So that's the step we're kind of at right now. When I say we're, it sounds weird. It's, like, myself and my robot are at Baird: You- you- your dev team of Claude subagents Brendan: Yeah. Yep Baird: Yeah. I, I totally get what you're saying though, and that's like with the agent stuff I've been building, it's kind of the… I'm do- I've done exactly what you're doing now, where it's like how do you get the more mechanical, getting the data in and start to figure out like what's the signal? You know, like what to pay attention to. But I agree, like the missing piece is how can it start act- and like this is a, a very… You've gotta be very careful about this, because like the next step is- Mm-hmm … how does it start acting for you? And so like I'm currently working on kind of the, the, that next step where it's like how can it start doing things like, not making changes for you, but like how can it start like writing the first bit, like, you know, a plan or whatever. Because th- that's always what I struggle with with trying to be productive, is that, okay, I've got a task, I know what to do, but taking that first step to doing it is always that like, you know, and we talked about this I think a couple ep- episodes ago, is like that activation energy of like I know I need to do it, I know what to do, but I can't get to it quite yet. Yeah. Or I'm not ready to get to it. And so like what is that like smallest thing- Yeah … the AI can do from the context it's getting out of your, you know, morning brief, you know, evening brief, whatever it is. To then start like working on your behalf to set it up so when it's like, "Okay, I need to pay attention to this," you've already got something to work from versus starting from scratch. Brendan: Yeah, exactly. And that, that's the other thing that suggested was, um, b- basically what you were just saying, like having things teed up in a way that it only really comes to me when it needs me, Baird: Yeah. Yep Brendan: been before, which is almost a summary of everything that has happened, no actual action taken to make it easier for me to get that thing solved. know? And, and I'm very hesitant and I'm not gonna give it access to, like, production for instance. So Baird: Perfect. Brendan: there are certainly queries and, uh, you know, requests that I get from people the organization that are asking for, you know, "Can you run this? Can you find this?" Whatever. I'm never gonna give it, you know, at least not write access to production, Baird: Yeah Brendan: there's something we get to where can have a read-only version of the database. I don't know. I'm not, Baird: Yep. Brendan: there yet. I'm not worried about that. Baird: Yep Brendan: security obviously is a huge thing, Baird: For sure Brendan: a- and I'm okay with that limitation, but what I want from it is more what you just said t- again, is like if you as the agent know that we're gonna need to run these queries, and you have enough context already because you have the code, the repo, to do the research. So, like, let's just take a, a not a real example, but a, a more concrete example. So my boss comes to me and says, "Hey Brendan, you know, I need a report on, you know, XYZ." That is something that the agent, because it's got the Slack message already, be able to have, I mean, as long as the ask has got enough context, which it usually does, should be able to then do the same thing that I would do today, which is, "Hey, I got a request from my boss, so, you know, now help me." What it should be able to do is already infer that, okay, Brendan got a request. How do I take the first part of it off his plate and tee it up for him to just, you know, knock it off the tee basically? So- Baird: No, for, for sure. And I mean, that's, you know, the agent stuff I've been doing, that's one of the rules that's literally in the agent's MD is, like, no write access. Like, essentially, like, unless I specifically create, you know, a tool that has write access for a certain purpose, everything's read-only. And it's, like, even baked in, like, when we're scoping things, it's always read access, you know, to make sure that it's like, even if it wants to write, it still can't. Because, I mean, these days you're probably gonna be good and 90% of the time it won't do something stupid, but you know, taking down production one time out of 100 is still gonna be pretty catastrophic, you know? So it's like trying to come up with the rules of, like, what does that actually look like is important. A- and I think another thing that's interesting to think about is, like, when, like… Because here's the problem I run into, and I think a lot of people will run into, is if you try to make a system like this be, like, a one size fits all, where it's gonna automatically be able to adapt to everything you need to do, you're gonna get into issues where it's like, oh, well, there's a gap there because it can't quite figure that out, or there's a context issue there. And what I would almost do, from your perspective, is, like, have Claude evaluate all of the things that you're doing, like, all the context it has, and say, like, "What are the things that we could actually set up?" So it's like, if you've gotten several re- requests for, like, a report to get pulled together. Like, how could you create an environment where it has the access it needs without giving it too much access? You know, maybe it's even, like, building a dashboard that runs locally on your machine that it's like you can pull that up, get it, export it, something like that. Like, you can start actually having it kinda categorize these different tasks it could take for you, and then you could figure out, like, what does that actually look like? And Brendan: Yeah Baird: a coding agent, it can build one-offs, it can build, like, little utilities for you that then it has access to run to kind of repeat these actions. A- and I think that's a really important kind of scoping of agents in my mind, is like, too much of it is, like, the, you know, open claw where it's like, it can do everything. It can learn for you. Like, that's cool, but in my mind, like, having it take a, a pass to figure out what you actually do, and then work together to start figure out, like, what are the actual right lanes that you go back and forth on, to give it the tools to then start acting in a responsible way. You know, in a way that you dictate versus it does, but it does the legwork of finding those patterns. Brendan: Right. Yeah, uh, that's definitely something that I've been thinking about. And, you know, you also just, as you were talking there, it just reminded me of one of the things that Jesse talked about, which was in the context of Grok Bot, and again, it could be anything, but Grok Bot works out well because it is, you know, one… I, I think each of the agents basically has its own computer, Baird: Oh, sure Brendan: have its own context and all those sorts of things. And of course, you can have them interact in the ways you need to. But the way he was treating it, and this is a little different than what I've done so far, you know, each one is like an employee. So need to find the things, like i- if, let's just pretend it was a person, like, you know, what are the things that you would give that person to be able to do their job? know, and then, you know, m- in some instances, you can give them every- you can give the agent everything. In other instances, you know… And but honestly, like, I don't really want… the only one on, on my team that can even go to production and, write things in the Rails console anyways. So that's already limited. I'm just a employee. No, I'm just kidding. But I'm, but I, I have the access because of, Baird: Yeah. Yep. Brendan: I… To your point, though, I would rather it tell me only I had access to this or we could do something else and use it, especially if, like, the re- same kind of request keeps coming in. And I've done a little bit of this too, where not necessarily through my, my agent, but just in general, like, my thinking is, okay, I've gotten this request four times. Time to just build it, and then I'll, you know… And, and that's a whole another thing that I don't even know that I wanna get into at the moment, but the coding side. I, I am doing a little bit more coding with the agents, that I don't want it… Like, uh, my, I don't let my Claude Command Center agent run that. becomes a separate thing, but then it can see the pull request, of course, and all that kind of stuff. But there's a wh- the whole workflow of handing off something like that. So, like, in this case, back to what we were just talking about, if the agent identifies something that you've gotten this request four times, and maybe we have a rule that is if we've done it four times, like, think about a way to, know, make this better or is there code we could write so it is deterministic at that point that Brendan doesn't need to build it. You know, no one needs to build it. Somebody else can access it or maybe the person requesting can access it. You know, things like that, that in, in this case, you know, obviously I'd be totally fine with an- it spinning up a separate agent or something or kicking off maybe a, a whole another coding agent that is, you know, the expert in application, something like that, instead of the, the more generalized command center thing. way, like that- That is another thing, like, we could do. I'm, I, I don't feel as strongly as a need there, Baird: Sure. Brendan: most of that also comes down to, like, and th- this is something that I don't, I don't think I'm realizing now, but just to kind of follow up on where we're all talking, or what we're all talking about here. You know, I think part of the reason why I feel this friction to begin with and why the command center thing I keep wanting to push on, or at least finding something else or something that'll do this, like, these are the parts of my job that I have to do, and it's not that I don't want to do them, but they're also the Baird: Yeah. Brendan: that, like, take my attention away from the things that are, like, essentially the Baird: it's it's context switching for necessity, but kind of brunt force, you know, Brendan: Yeah Baird: ver- like, it's not the high level thinking. A- and especially in your position, it's like this is the type of stuff that having something or somebody take care of for you is valuable. And, you know, I, I think you, you're to the right point where it's like if you have to keep doing it, it's probably automatable, and then it's worth actually looking into Brendan: Yeah. Yep, exactly. So yeah, I mean, that's, that's where I'm at in the journey. I, I definitely getting to a point where I can report back on more and, you know, I, I don't think I need something like Grok Bot anyways, like at least not with this role. And, and, you know, in Jesse's case, he's running Bento, and he's using it for all these things like support and stuff that, you know, we have people doing that. I don't need to do that myself. So I, I don't feel the need to completely knock out all those things and like… Or I should say, like use like Grok Bot necessarily. So I'm hoping anyways to just sum up where we are at the moment, you know. It's like I'm hoping that this, uh, ba- basically like the things that I need, I can get out of this system without having to go to something like Grok Bot. A- and even Zapier based on that, that interview, you know, they've got something that I don't, I don't know if I wanna compare it to Grok Bot, I don't… Didn't look into Baird: Sure. Brendan: enough. Baird: Yep Brendan: you know, the, the same idea of like an agent having its own computer and can do, know, more things for you, sorta like Grok Bot in the way it's been described is there. So the, like there's options out there, right? There's lots of them, more Baird: For sure. Brendan: too, but Baird: Yeah, and I think the, the last kind of bigger thing I'll say is, and we talked about this when we were just talking offline last week. The thing you gotta be careful with, with all of this, when you start handing the AI the keys to start doing things for you, is then everything's a layer, layer of abs- as- abstraction away from your thinking. And so, and this is like when we've talked about building things. It's like when it does all the planning for you and you're not, like, connected to it, you start to lose context on your side of things. It's like, "Why did we do that?" Like, "Why is it this way?" And that's a really dangerous place to get. So it's like you gotta be able to be productive without shipping all of the cognitive load off of you, otherwise when it comes back around, you're gonna get caught out, you know? And, like, Brendan: Right Baird: it will be infrequent, but when it happens, it's usually not great. And, and so it's like, you know, we, we've talked about this a lot. It's like how do you find that balance of, like, you're involved, but only when you really need to be, but still enough that you know exactly why everything's happening? Brendan: Yeah. Uh, that's a really good point. And just to end, I, I think… The, the way I like to think about it is just delegation, right? It's like if the company had more money or funding or whatever, like I could hire some more roles that would fill some of these gaps. And so I, I definitely agree with that. Like, that, that is a major problem if you get to the point where you're relying on the agents to do everything, and you don't have any context. that's definitely not where I wanna be. But I want it, I want it to be something that I can essentially just delegate off to some- someone or something the job done, but then report back, and I need to be in that loop still because… A- and I, I also don't really like… I, I, I shouldn't say I don't like. I, I have not gotten to the point where I have authorized or even want my agents responding to people. Baird: Yeah, for sure. Brendan: one responding. I just need to get Baird: Yep. Brendan: to do that Baird: Yeah. I mean, you, you could get to a point someday with, like, the reporting stuff where you could plug into Slack and have your bot respond with a report to a Slack request, but that's, like, pretty straightforward, like, Brendan: Yeah Baird: versus truly communicating. And so I agree, like, having agents truly do the communication for you is a, a slippery slope, like, both from what they do, but also the trust that erodes from people that are like, "Why am I just talking to your bot all the time?" You know, like, it's kind of a really weird… And, and, you know, it'll change, but right now I just know there's a lot of negative feelings towards people responding through AI versus that human contact. So I, I think you've got the right thing there where it's like AI helps you get all the pieces together and do some of the execution, but it's still, like, you're driving it Brendan: Yep. Cool. All right. Well, um, yeah, thanks for watching and listening and, uh, know, feel free to rate us review us on Apple Podcasts, wherever you get your podcasts. Uh, you know, like and subscribe down below on YouTube, uh, we'll catch you next time