An open-source design for an AI agent that can outsource coding tasks to other agents, tools, and humans when it makes sense.
This repository is intentionally design-first. It describes the control loops, routing logic, interfaces, and safety boundaries for a system that decomposes software work and delegates pieces to specialized workers.
- A practical architecture for task outsourcing in coding workflows
- Delegation policies for deciding when to use another agent
- Contracts for subagents, reviewers, and tool executors
- Safety, traceability, and rollback requirements
- An implementation roadmap for a real system
The main agent should not try to do everything itself. Instead, it should:
- Inspect the task
- Break it into isolated subtasks
- Decide which subtasks can be outsourced
- Dispatch work to specialized agents or tools
- Verify outputs against tests and constraints
- Merge only validated changes
- Reduce context overload on the primary agent
- Increase throughput on parallelizable work
- Keep human oversight on risky actions
- Make every delegated action auditable
- Support multiple backends for subagents and tools
- docs/architecture.md — system design and control flow
- docs/delegation-policy.md — rules for what gets outsourced
- docs/worker-contract.md — input/output contract for subagents
- docs/safety-and-trust.md — security and review boundaries
- docs/roadmap.md — staged implementation plan
MIT