From One Agent to Many

Isolation before concurrency

Two agents working in the same checkout are not two independent workers. Their file changes, generated artifacts, processes, and assumptions can collide. The workspace root is the interactive driver’s checkout; it is not a shared mutation checkout for bounded Runs.

Give each active Run its own execution environment. A Git worktree is the default reference implementation, but Docker, devcontainers, VMs, or another provider can implement the same isolation principle.

workspace control repo / driver
     │
     └── repos/api (canonical managed repo)
             ├── Run A / worktree-a / feature
             ├── Run B / worktree-b / bugfix
             └── Run C / worktree-c / housekeeping

Coordination is separate

Isolation prevents routine file collisions. It does not decide whether two changes are compatible. The driver shares minimal coordination information— usually Intent and status—while Run plans remain private by default. Generated AGENTS.md and .agents/skills/ tell the driver how to maintain this split.

The integration seam

Each Run produces a Candidate. Validation and integration happen after execution, with the repository’s normal review and CI mechanisms still in charge.

This is the point at which the architecture becomes a system of bounded Runs rather than a collection of immortal agent processes.

Continue with Multiple Agents / Same Repository.