00:00:00Imagine if Clod Code wasn't just one assistant, but an entire team, 20, 30, or even 50 agents
00:00:06working in parallel, shipping features while you watch.
00:00:09This is Gastown.
00:00:11It's not chat with an AI, it's a factory.
00:00:14It decomposes features, assigns parallel workers, merges code automatically, persists everything
00:00:20in Git, and keeps running even if it crashes.
00:00:23This could 10x your output or just cause things to go south.
00:00:27Let's get into the setup and how this all actually works.
00:00:35Now Gastown is an open source orchestrator that turns AI coding agents into something
00:00:39closer to Kubernetes for dev workflows.
00:00:43First thing, Gastown is not an AI model, it's just an orchestrator.
00:00:48It sits on top of tools like Clod Code and turns them into coordinated multi-agent systems.
00:00:53You give a high level goal to what they call the Mayor, and the Mayor breaks that goal
00:00:58into structured task units called Beads.
00:01:02Beads then get grouped into what they're referring to as Convoys, and then the worker agents called
00:01:07Polecats execute them in parallel.
00:01:09Now this is huge for a few reasons.
00:01:11Everything runs through Git.
00:01:13Every agent works on branches, merges queues, state is persisted, so if it crashes, you don't
00:01:20need to start all over again.
00:01:21Now compare that to normal vibe coding, right, just AI coding in general.
00:01:26We have one session, linear execution, context drift, manual orchestration, it's kind of a
00:01:31mess.
00:01:32Gastown is basically horizontal scaling for cognition, and that's a big deal.
00:01:37But the real question is, does this actually work?
00:01:40If you enjoy this type of content, be sure to subscribe.
00:01:42We have videos coming out all the time.
00:01:44Now let's get in, fire it up, and do some tests.
00:01:47Alright, so I've got a simple to-do app running.
00:01:49I found a super simple Go-based one on GitHub.
00:01:53I chose Go because Gastown is actually mainly written in Go, but really there's no difference
00:01:58in the project you choose, it'll still execute like normal.
00:02:02Now this project, this to-do app, there's no auth, it's just a basic CRUD application.
00:02:06And if you've ever added auth into an existing app, you already know the pain of breaking
00:02:11things.
00:02:12It's never just add login, it's database changes, routes, middleware, forms, tests, docker, everything.
00:02:18And when you're building these apps, everything is touching everything.
00:02:22The challenge we are testing is, can Gastown do that without me micromanaging it?
00:02:26Now honestly, the setup here was kind of a pain.
00:02:30I needed to brew install a lot of larger packages, I had to get everything in order, and getting
00:02:35it running once it was even installed took some troubleshooting, it took some time.
00:02:39But once I managed to sync everything, it does run really well.
00:02:43Now that it's running, I attach the mayor and I give it one prompt.
00:02:46I'm going to say something like add JWT auth, use SQLite, it's basic, create login and register
00:02:52endpoints, protect routes, add HTML forms, write tests, add a docker file.
00:02:58That's it.
00:02:59That's a lot of stuff though, right?
00:03:00All these things going on.
00:03:01No step-by-step prompting, and let's watch what actually happens here.
00:03:05The mayor now decomposes these features automatically.
00:03:10The schema, middleware, auth endpoints, frontend forms, tests, those become separate beads,
00:03:18and remember, a bead is kind of like one of these workers.
00:03:21And now look, all these polecats are now swarming.
00:03:25One agent creates the SQLite schema, one's going to write the JWT middleware, one builds
00:03:31login and register routes, another updates the UI forms, and then one writes integration
00:03:37tests, our docker, all this is happening at the same time.
00:03:40And here's really the big contrast.
00:03:42If you were using Clawed code alone, you would do this sequentially.
00:03:47You'd prompt, wait, fix, re-prompt, re-explain context over and over again, right?
00:03:52That's kind of how these code agents work.
00:03:55But here Git branches isolate the work, merges are queued, state is tracked.
00:04:00So instead of you holding the whole plan in your head, the system actually holds it.
00:04:05And if this crashes halfway through, that's okay actually.
00:04:09You resume from persisted state at a later point automatically.
00:04:13Now the moment of truth here.
00:04:14I forked this into a repo, so now I want to go through and look at all the changes it made
00:04:19and added right here just on my GitHub.
00:04:22We can see all the old code, but then everything else here was edited.
00:04:25So register, login, token issued, protected route works, tests pass, all this stuff.
00:04:32This is what was added in through Gastown.
00:04:35That entire feature decomposed and implemented in just a few minutes.
00:04:39And that's the point.
00:04:40The difference isn't just speed, it's cognitive offloading.
00:04:44We are no longer managing every little AI step, but that doesn't mean it's automatically a
00:04:49win.
00:04:50Because running this could be another story.
00:04:52Now let me run this real quick.
00:04:53Let's kind of see what we get and what we're working with here.
00:04:56So I mean, all in all, this is what it looks like.
00:04:58It's not too bad, super basic app.
00:05:00And you can see that I can log in here, create an account.
00:05:03The to-do list still functions, it didn't really change anything with that.
00:05:06It added our routes, the Docker files, all that extra stuff with JWT authentication.
00:05:11So it did really good here.
00:05:12On a larger scale app, I would love to see how this run, but this really ate through a
00:05:16lot of my tokens.
00:05:18Speaking of that, all right, now let's talk the trade-offs.
00:05:21So pros and cons, right?
00:05:23First the pros, scalability, right?
00:05:25This was really cool.
00:05:27You can have 20 to 30 agents working at the same time.
00:05:30That means massive output on big features, the kind where one person gets buried in tasks
00:05:35and interdependencies.
00:05:37Second, another really cool feature here is the Git persistence.
00:05:40No more fragile AI sessions where everything vanishes if something crashes and we have to
00:05:45pick back up.
00:05:46Right?
00:05:47Those are really, really cool things.
00:05:48But now the cons here, because this is where people get surprised.
00:05:50I was also surprised at some of these things, but they make sense.
00:05:54Oversight fatigue because yeah, the system is producing a ton, but you're also now reviewing
00:06:00a ton.
00:06:01It can be super productive, but it's also going to be super draining.
00:06:04Then we have the cost.
00:06:06Heavy usage can get expensive fast.
00:06:07We're not just running one agent.
00:06:09We're running all these agents in parallel.
00:06:11I'll let you do the math on that.
00:06:13Agents can still conflict.
00:06:14They can over engineer.
00:06:15They can absolutely wreck your repo because it's giving them full reins on everything.
00:06:19This isn't just plug and play, which leads to the real question.
00:06:23Would you actually want to live like this using something like this in your workflow or is
00:06:27it almost too much?
00:06:29Yes, you can choose other providers too, though the choice is yours, Claude, OpenAI, okay.
00:06:34So yeah, this is an early stage tool, but it's pointing towards something so much bigger
00:06:38because the direction is clear that this isn't AI helps you code.
00:06:43This is AI changes the workflow with all these agents running in parallel.
00:06:47So Gastown represents this shift from AI assistant to lots of agents now working together.
00:06:54This changes our leverage because now you can orchestrate backend, frontend, tests, docs
00:07:00and infrastructure all at the same time.
00:07:03This could make you dramatically more productive or it could redefine what development even
00:07:08means.
00:07:09The question isn't whether multi-agent systems are coming.
00:07:12They're here, right?
00:07:13This is an example of that.
00:07:14The question now becomes, are you going to experiment with it now or later?
00:07:19Because who knows?
00:07:20They're coming.
00:07:21Subscribe for more tools and tips in the tools reshaping dev workflows.
00:07:25We'll see you in another video.