September 30, 2026
Capy AI agents split the work themselves: a simple prompt can handle coordination
Capy launches child agents and passes messages between them when asked to split the work and coordinate changes.
Garry Tan
@garrytan
Amazing: you can simply ask Capy's coding agents to coordinate their work and avoid editing the same sections of code as much, and Capy adds communication between agents itself, with everything working without any extra coordination setup
· 7.3K views
In an example in Capy's docs as of Sep 30, 2026, an agent splits an Express → Fastify migration into 3 parallel tasks. Each ends with a separate PR and tests.
Instead of configuring a separate orchestrator, the developer describes how to divide the work in a prompt. The AI agent creates, launches and coordinates child agents itself. Capy claims support for 25+ agents working simultaneously in one repository, with coordination across branches.
How to split the work. In the migration example, the agent assigns different parts of the application to different tasks:
- routes/auth; - routes/billing; - routes/webhooks.
Each task gets a fresh machine, a separate machine with its own copy of the repository. The docs recommend a shared machine for reading, auditing and analysis. Child tasks can create tasks of their own, nested up to 3 levels deep.
When tasks change different files, they run in parallel. If changes overlap, the docs recommend stacked tasks: the second task waits for the first to finish and starts from its PR branch. The agent can also create sibling threads, send them messages and explicitly read their status to coordinate. For routine parallel work, Capy recommends tasks.
How to get started. Capy runs in the browser at capy.ai. To get started:
1. Sign in at capy.ai and install the Capy GitHub App. 2. Grant access to selected repositories and add them to a project. 3. Choose a model and send your first prompt in a thread.
You can use the prompt to assign an Express → Fastify migration, list the three parts of the application and request a separate machine, PR and tests for each task. As of Sep 30, 2026, desktop app version 0.3.0 is available for macOS on Apple Silicon and Intel, Windows x64 and Linux x64/arm64.
Capy plans as of Sep 30, 2026:
| Plan | Price | Credits | | --- | --- | --- | | Lite | $1 for the first 7 days with a verified card, then $20/month | $20 on the monthly plan | | Pro | $100/month | $105 | | Max | $200–$1,000/month | 10% bonus |
Capy lists unlimited concurrent threads. Credits cover the chosen model's tokens and machine time: a standard Large machine costs $0.40/hour, so parallel tasks also incur charges for each machine's runtime.
A sleeping machine is free, but Capy bills for another 2 minutes of runtime after it goes to sleep.
Source
