Claude Cowork Scheduling & Permissions: Fixing Friction in Autonomous AI Workflows

Brendan

I, I think I've been kind of moving away from it a little bit. I, I still use it for some things. Um, I, I've run into some issues myself recently. And it sounds like maybe you've run into some of the same ones. So. Why don't you kind of talk about what you've run into and, and I can let you know if it's the same thing or if I've got different stuff that I'm seeing.

Baird

Yeah, I mean, it, it is something where I keep wanting to use it and I keep pushing on it to see how far I can take it. And there's a couple things that I've run into. You know, one of them is just that two, get it to work. When you're scheduling things, you know, your computer's gotta be on, which it, it makes sense, but it be, can be kind of a source of friction because sometimes like. If you're running it in the morning and you're a little bit late getting to your computer that day, like it skips it. Now, if you go back in and look at it, it'll like pick it up and run it. But it's like there's a little less consistency there, which, sure, I could turn my computer on 24 7, but for what I'm doing with it, it just feels a little like overkill and it feels like there might be other better things. But I think the, the biggest source of friction I've seen, and again, like we've talked about this off and on.

Probably in a lot of the episodes that we've done so far is like, just around permissions and like, it seems like you'll give things permission, you know, to go out on the web, grab things or work on files, but then all of a sudden, you know, a couple runs later, it runs into another issue where it just won't do it and you know, it, it doesn't seem to fail. Loudly enough where it's like you necessarily notice. 'cause at least for me, it's like background processes. It's just things that are happening that I want to happen, but they're not in my face enough for me to always notice. And so I, I've definitely just run into more issues with. You know, permissions in general. And I've noticed too that Claude Code as well, permissions have kind of gone up though. I know that they're doing a lot of work with like auto mode that's just come out and stuff like that where they're trying to address it, but it just feels like a bit of a, a messy time right now on that side of things, what have you been seeing?

Brendan

Yeah, I mean, I, I've been seeing a lot of that too. And, and I think for those that maybe didn't listen to our last episode, you know, we talked a little bit about YOLO mode or what we call YOLO mode, but know, in permissions and stuff and, and in the context when we talked about it last time, you know, it was about, uh, you know, I, I was talking about the productivity tool that I was using with Claude. Primarily in the terminal. I mean, I was, I do kind of integrate Claude Cowork with it a little bit. Um, but I, I've pretty much seen the same thing that, that you were seeing and kind of a funny story. So I, I set up Claude Cowork, I guess it's been at least a month ago, I guess now. I, I kind of, when it came out, I set it up. I have, I don't know, like six or seven scheduled jobs that I have run every day. And you know, to your point.

You have to have the laptop open. And so if you're using a laptop like, you know, both of us do, and you don't have a Mac Mini or whatever running in the background, um, you know, you're, you're at the, the whims of when do you have that open for it to run. And that's one thing. But even when I have had it open, my experience in within the last week has been. I kinda lost track of it and I didn't realize, 'cause you know, mine, I have it set up so I get notifications, I just know, okay, go check out that this job is done. Um, whatever reason, I, I hadn't seen any of those notifications.

didn't even check myself to go look. What, what I ended up finding is that I had. I think there were, I think it shows like five or six of like the, the most recently run jobs.

Baird

Yep.

Brendan

have more that run, you know it's not gonna show everything. of those, you know, every single one of them had a little blue dot or whatever next to it showing indicating that it was ready. I opened each one. And I think those six, I think it was, I think all or five of the six at least, was waiting on me to give it permissions like you just said. And. Some of them too though, because I, I had, uh, I think like two of them where I hadn't even looked at the output for like a week. And so what I also noticed is that a, a couple of those, the last time they ran was like three days ago,

Baird

Sure.

Brendan

it wasn't even that they had run today and they were

Baird

Yeah.

Brendan

They had just. stopped running, and I don't know if it's because they were looking for

Baird

Yeah.

Brendan

I, I did notice when I like allowed a lot of it and, you know, basically got caught up. It seems like it's kicked in and it's

Baird

Sure.

Brendan

But that was kind of an interesting thing. I don't know if you've seen that at all either, but, yeah, the allowing permissions is kind of the, the interesting one because it, it gets annoying when you have a scheduled job that you've already done multiple times and you want it to, to just run. But then it sits there and it waits for you, and that's part of what caused my problem. But I, I cut you off there.

Baird

Oh no, I, I think it actually speaks to kind of a bigger thing in my mind, like just in general is like coworks cool and like I use obviously Claude Code, I use Codex. The thing that I kinda struggle with is that they're all kind of separate. That they all have different tabs. They've all got different projects. Like you've got so much spreading out now where like, especially, and I've heard that, you know, OpenAI is now planning on combining Codex and Atlas and ChatGPT all into one app, kinda like Claude has done or Anthropics done with Claude. So it just feels like. Well that's kind of cool 'cause then you don't need three different products to do, you know, the different things. There's like a UI layer that's missing, that's like, how does this all come together? Especially if you're running autonomous, you know, processes. Where's the dashboard that actually like tells you everything you need to actually pay attention to? And like maybe that's something where when it feels like a missed opportunity from.

You know these huge companies that this is their stuff, but it also like maybe there's an opportunity to tap into that in a different way, but it's just something I've noticed to your point of like when something stuck, yes, you can look in the sidebar in cowork if you think to go there to look for it, but how can that get surfaced in a more interesting way?

Brendan

Yeah, and I mean, you're also making me think like neither one of us again has used dispatch,

Baird

Yeah.

Brendan

a, a layer on top of cowork. And again, from my understanding, it's basically just a way to, to make it more like open claw. You

Baird

Yep.

Brendan

to it when you're not at your desk or even at your house, whatever. you have to have your computer running.

Baird

Yep.

Brendan

So that's a limitation. But you know what, what it made me think about is like if we were running dispatch even, like would that, even if you're sitting at the same, right in front of your computer, but you get dispatch and it's like maybe pinging your phone. don't know if that would be easier or like a better way to handle it, or at least a way to just like keep on top of it. know, I'm maybe, I don't know.

Baird

Yeah, I mean, it, it is definitely something that would be, I have to look at it 'cause I had actually did set up dispatch on my phone. So like I've got the app on my phone and I don't see anything right now. And I don't know if like, I'm just missing a syncing step in there or something, but at least just like out of the box after you set it up too. Work with your computer. It doesn't just do that automatically. So that might be something that's worth investigating for us to see. Like is there something there that can help us?

Brendan

Yeah, think so. I, I think maybe that's something that would make sense for us to report back on next week is, you know, why don't we just take some time, mess around with dispatch, see if that helps with some of these issues, know, but, but at the same time, it's not gonna, that, that's not gonna be the solve for the, you know, uh. Just allowing it to do things or whatever. Um, so the permission stuff, so. One thing that, you know, we talked about before we started recording as well, is what if we just in May, maybe the,

Baird

Yep.

Brendan

and tell today is, let's actually go in and talk to Claude live here. don't know if we're gonna get to a resolution, but let's see what Claude can do to help us if, even if it's not in cowork. You know, I don't know if, I sometimes wonder too, this is kind of a different thing, like when I'm running. Claude in different places, different harnesses or wherever. I will ask questions of Claude, about Claude. And half the time it's like, I don't know anything about that,

Baird

Yeah.

Brendan

if the models are updated to know enough about cowork, but it's worth a shot. I'm kind of interested to see. Um.

Baird

Yeah, and I mean, honestly, while you're starting to spin that up, this actually touches on something that I've found in a lot of the work I've been doing recently where one of my favorite things about working with AI is that you can actually talk to your AI to figure out what it does know and what it doesn't know. Like I've many times with the product I'm building. For work. It's like I'm trying to figure out, do you have access to the tools that I need you to? Like what data are you seeing? You can actually ask it and it'll come back and talk to you about it and tell you what it knows what it sees, and it's like having a interactive debugging partner essentially to truly like, because it's a challenge I've always faced is like. There is a bit of a black box, even when you're like putting pooling and a harness around the AI of like, what is it actually seeing and what is it actually doing? And it's kind of fun just being able to interrogate the thing you're working with to see what it actually knows and see how it can help you get to the right spot.

Brendan

Yeah. Yeah, that's a good point. Let's go back to the cowork example, and I think let's just give it a shot and say, um, Claude Cowork, I'm not gonna say everything out loud, but. Uh, my cloud coworker is asking for permissions than it seems like it should on scheduled jobs or something like that. Can you help me?

Baird

It's actually is an interesting thing that I'm seeing though, in that when I was in the desktop app and I was doing something similar to this, it actually didn't have permissions, I don't think. As a slash command.

Brendan

Yeah. Well, I. I kind of was wondering about that too. 'cause you seem surprised by it and I figured you would've tested it to or checked it out too. I've noticed that sometimes too. I don't know if it's 'cause like I, I did just, although this is interesting, I, I did right before this, I just refreshed or relaunched my Claude desktop app.

Baird

Yeah.

Brendan

So that's a fresh version. I don't know if that would interact with my Claude Code instance here or not. I don't, I would be surprised if it did, honestly. Let's see. It's prompted me for something. Use skill update config. Okay, look at that. it's basically telling me you can use update config as a a skill, and then it's got instructions basically for how to do it. and it's also asking me permission, do I wanna proceed?

I guess I'm trying to decide like if, if I say yes. So the, what it's basically saying is, do I want to update the permissions on update config without asking again? So here we go. It's asking me for permissions. Um,

Baird

I'm not sure.

Brendan

gonna do yet. So let's see. It's saying, use this skill to configure Claude co, the Claude Code harness via settings.json. So I, what I'm seeing here is that you can. Potentially set up in Claude Code, something that would work for any harness and that it would include Claude desktop.

Baird

Yeah, because I, I think there's just one shared, 'cause like in your directory you've got the dot claude folder, and under that there is the settings.json. So. I think, I mean, I, I haven't dug into it too much recently, but I think whatever you set there actually works for every different, you know, essentially type of club that you can work with. Lemme look at one of mine.

Brendan

That it would make sense. I guess for me, I haven't taken the time to dig into where or how Claude Desktop works. I know it's essentially a layer on, it's, it's a har, well, it's a harness for Claude Code,

Baird

Yeah.

Brendan

but I guess what I mean by how it works is I don't know that I've ever taken a look to see. Like what, what Claude Code is it using? Is it coming with its own version in like, say a, a VM or something? Or is it using, you know, the same version that I've got on the, the desk or on, on my, uh, terminal.

Baird

That's actually a good point because I've run into an issue before where. Back when I used to use the plugin in windsurf for Claude Code, I would run into an issue where it wasn't working in that environment, but if I went to the terminal directly, it worked perfectly fine. And so I think there might be. Maybe not complete differences, but it does seem like they're using a little bit different versions depending on where you're at. And I'm sure when you're in the desktop app, it's again, another environment because that one's running inside of the desktop app versus like in your terminal or in windsurf. So there's definitely, I think, some nuance there.

Brendan

Yeah, yeah, for sure. And you know, just coming back to what, what we're doing here in the terminal again, like I, went ahead before and I said. Well, I guess my prompt's gone, but now, but you know, it was basically saying, do I want to update the config, um, settings? Basically I said, yes, allow it. Now here's a good. Example of why it's good to read what Claude is telling you or Codex or anything, I'm realizing now what it's trying to do is it's, it's telling me the fix is straightforward. you have to do is set default mode to bypass permissions so that we will never, basically, it's never gonna ask for permissions again. And that's not what I want though,

Baird

Yeah.

Brendan

right? Like, I'm not gonna say yes to this because I don't want it to say that. 'cause I don't, I, I don't want to do that because it, who knows what the heck it's gonna do.

Baird

It it, it is interesting though, 'cause if you read what it's saying, it says for scheduled or automated agents, you need to set default mode. But if I'm looking at the config. How, like where it's editing, it doesn't look like there's any, you know, agent specific setting there. It feels like it's just setting permissions globally to bypass as default.

Brendan

Yeah, exactly. I'm gonna say no. And you know, I, I think, again, I, I wasn't too concerned if we actually solved this, but I wanted to just show what it's like to dig in a little bit and, and talk to Claude about this

Baird

Yeah.

Brendan

I am gonna push back one time here though, and I'm, what I'm gonna do is. And this is something I like to do a lot, if it has a specific thing that it says, I don't really wanna type it out or whatever, um, but I will say, I'm just gonna even copy the whole thing this time and then ask it a question. Now, granted, that might already be in the context, but. For me, I just sometimes like to be very explicit

Baird

Yep,

Brendan

don't, don't, I don't want to let it guess

Baird

for sure. Yep.

Brendan

at. anyways, I'm gonna just gonna say, okay, you told me, or actually I wanna say it appeared that you were going to basically give permissions at a global level. Um, you know, based on the config file we saw and. I, I just want to question it and say, am I, am I missing something or did you, you know, tell it to me wrong?

Baird

No, I, I definitely think that that's helpful. I almost daily, you know, probably several times a day essentially have to ask questions like this, where it's like. Are we understanding each other or you know, is there something off and, and I think it's a good gut check. You know, it's like, you might be wrong or we might be wrong in this case, but if you just assume that the AI knows what it's doing, you can get yourself burned pretty quickly.

Brendan

Yep. here we go. The, the answer it gives back is, you're right to catch that. Bypass permissions globally means no permissions, prompts anywhere. Interactive sessions two. That is too broad. So the honest answer is Claude doesn't have a way to set permissions only for scheduled jobs versus interactive sessions through settings. Do JSON. It's the same file for both. So that's interesting.

Baird

Yep. Yep.

Brendan

I, I'm also wondering too, because I know that Claude, like if we just do slash schedule Yeah. There's, there's a scheduler. In Claude Code here. And I also wonder too if like, yeah, I asked it about Cowork, but if it was thinking, Hey, schedule jobs inside of me, Claude Code here, not co, uh, cowork. and that, that's something that like, and. It could have been a bad prompt. I'm sure if I were to ask it, it would be like, oh no, you were fine.

Baird

Yeah.

Brendan

It wasn't you,

Baird

Yeah. Yep.

Brendan

but it again, it's like, yeah, like right here, it's telling me that bypass permissions is the nuclear option. Fine for a personal command center where you trust everything. But it's, it, you know, because it's a global setting, we don't wanna just do it for everything.

Baird

Yeah.

Brendan

Um, you know, I think. You know, we don't have a whole lot more time to record here, so we're, we're not gonna get to a resolution anyways, but it, it does kind of make me want to keep digging into this and see if there is a way to figure out, more granular, like, okay,

Baird

Yeah.

Brendan

want to accept the things that we were talking about and. May, maybe that's something to just do a little more research and we can, uh, you know, report back on here. But I think I'm just gonna leave this where it is. I'll just basically end the session.

Baird

Mean, I think it's a good experiment and I think. It feels like a hole. You know, with, with them pushing more cowork in Claude Code, it feels like there needs to be some separation of the config and maybe there is like it. It definitely something that we should keep researching because it just feels like there needs to be a way to do it. And I do know. I think the other way around this is that you can be a little more specific in general where it's like, I'm cool with web searches and I think there might even be a way to like whitelist what it can, web search and stuff like that. So, you know, I, I think that's the next iteration of this is no, there's not a safe nuclear option, but what's the in between? And I think another thing we could try maybe next time, or we could at least report back on is like. Doing an interactive session with Claude, 'cause I know it can edit your settings. Do JSON like, so you can kind of talk about like, what do I actually do frequently? Where are the hangups and you know, have it, 'cause you know, the thing I've noticed is you have to know like all the different commands the right way. Like, it's not like, of course, you know, an LS is, but like.

Sometimes you have to do like the LSR, sometimes you want it to be more, you know, fine tuned than that. And so you can talk to Claude and have it do the edits 'cause it knows what it needs. And then you can kind of interactively find that kinda white list of what is like the sane bypass permissions that you wanna put in place.

Brendan

Yeah. I think that's a good point. 'cause like one of the examples I believe I brought up last time we recorded was if I'm gonna run a Python script. You know, probably okay with a specific script that it's gonna run, but the command that Claude sees that it's going to allow, at least based on the way it would've done through the prompts we just saw there. it's going to

Baird

Yep.

Brendan

all the parameters that make it unique. You know, especially if it's say like an API call, that's calling a or being if, if the API call to, like, let's just say we're we wanna go to google.com for some API call, and then the next time we want to go to facebook.com or whatever it is. But you're basically calling the same script every time and it's only the difference of Google and facebook.com. make Claude think it's a new thing that it like will continually ask for.

Baird

Yep.

Brendan

the sort of thing that I was getting at last time. And yeah, so I think some way to whitelist, you know, and maybe it's the same kind of thing that that LS Star

Baird

Yeah,

Brendan

like,

Baird

I think it probably is, but, but you, you don't wanna do Python star 'cause then it can just,

Brendan

Python, it's gotta be your script

Baird

So, so it can't, and that's kinda what I'm saying is like finding, uh. That in between. 'cause you don't want it to ask for permission for all of the different parameters, but you don't want to give it permission to do any Python execution because then you're opening yourself up for all sorts of problems. And so

Brendan

Yeah.

Baird

I think the interactive way of doing it could be an interesting way to just like, and some of it too is, I'd be curious to see, and, and again, this can be a follow up, is like. Can Claude look into your like command history even, because like you can, you know, on the terminal, look at what commands you run and can it even look through there and be like, okay, this is what you do. Run and see, you know, and use that as like a, a baseline. And it's always, how can you use AI to make the boring but important stuff easier? Because it, you know, if you have to do it all by hand. You're less likely to do it. You're either gonna go with the bypass, all permissions, or you're gonna be hitting except all the time. And so like, where's the best way that you can do a more robust version of it without having to just like go through and do it all manually.

Brendan

Yeah. Yeah, exactly. Well, I don't have a whole lot else to add to that. Um. Yeah, so can leave it there, but I think there's definitely things to check out and report back on. Um, you know, like we said, into dispatch. I think. We'll, it'll be interesting to see kinda, if that least improves some of the, the scheduling workflow stuff that we talked about with, uh, cowork. So, well, on that note, um, you know, feel free to. Check us out with our newsletter. Um, that's live. We get probably by the time this is live, we'll have a couple, uh, issues out. Um, you know, feel free to like, and subscribe down below. And, uh, yeah, looking forward to the next one. you.