LARPCO is an open-source autonomous agent-company framework that runs inside GitHub Actions. It spawns a full cast of named AI agents who plan work, debate approaches, seek approvals, and ship code.
Corporate cosplay that actually ships. The PRs still merge.
LARPCO spawns a full cast of AI agents inside your GitHub repo. They have roles, titles, and a chain of command. Under the costume, the architecture is solid. The bit just makes it more fun to run.
Orchestrators, PMs, engineers, reviewers, scouts. They plan sprints, write PRDs, debate approaches, open PRs, and post standups to Slack. You're the CEO. That's your entire job.
Every engineer agent runs in a fresh context window with exactly one task. One file. Ships it. Goes home. The agent that writes code never reviews it. Separation of concerns, taken literally.
Every completed loop writes its learnings back to a shared ruleset. Bugs get documented so they can't recur. The system gets smarter with every task it ships. Dot remembers everything.
Approval checkpoints via Slack. You approve scope, not implementation. A reaction emoji kicks off the pipeline. You get a ping when something needs a human. That's it.
Alma catches bad plans before any engineer spends tokens on them. A bad loop plan wastes every token downstream. Plans are checked before workers run. Always.
No proprietary platform. No vendor lock-in. The pipeline runs on standard CI infrastructure. One script. Auditable. Forkable. Already works.
Every LARPCO instance ships with a named company of agents. They have roles. They have vibes. They don't have personalities — they have a chain of command.
| Name | Role | Vibe |
|---|---|---|
| Dottie | Orchestrator | Runs the whole show. Has seen everything. |
| Xanthe | PM Agent | Very into process. Writes long PRDs nobody asked for. |
| Skye | Design Agent | Sends mood boards. Cc's everyone. |
| Blaze | Eng Lead | Decomposes your tasks. Very confident about it. |
| Alma | Loop Reviewer | Catches everything. Trusts no one. |
| Bindi | Engineer Worker | One file. Ships it. Goes home. |
| Velma | QA Reviewer | Finds the bug. Every time. Not sorry. |
| Pixie | Feature Scout | Fires after every deploy with unsolicited ideas. |
| Dot | Learning Agent | Quietly updates the memory. Never speaks otherwise. |
A task enters as a Slack approval and exits as a merged PR with passing tests. You approve the what. The company handles the how.
GSD says "no enterprise theatre." LARPCO says "we exclusively do enterprise theatre." The PRs still merge.
| GSD | LARPCO | |
|---|---|---|
| Vibe | Minimalist, efficient | Corporate cosplay, chaotic good |
| Structure | Commands you run | Company that runs itself |
| Human role | Developer | CEO |
| Agents | Subagents, anonymous | Named cast, roles, chain of command |
| State | .planning/ directory | ORCHESTRATOR_STATE.md + git |
| Approvals | Interactive prompts | Slack reactions |
| Tone | "No bs" | "The bs is the feature" |
| Works? | Yes | Also yes |
Every decision in the system was made to prevent a specific failure mode. The comedy is in the framing, not the output.
Agents accumulate bias and drift when they carry context across unrelated tasks. Every worker spawns with a clean window and exactly the context it needs.
An agent cannot objectively review its own output. Velma and Alma are always separate instances from the engineer. Non-negotiable.
Xanthe writes better PRDs when she's never touched code. Blaze plans better when he doesn't also have to implement. Specialization compounds.
A bad loop plan wastes every token downstream. Alma exists to catch under-specified plans before any engineer spends tokens on them.
Dot updates the shared ruleset after every completed loop. Bugs discovered in production get documented so they cannot recur. The system improves with every task.
You decide what to build and approve task scope. The agents decide how to build it. This keeps you strategic and the pipeline fast.