Async Agent Workflow

Goal

Async agents should make Prox OS faster without turning the repository into a shared edit buffer. Parallelism is useful only when each worker has hard scope, separate files, validation, and an integration path.

Standard Workflow

Use this workflow only when the user explicitly requests async agents, multi-agent work, task graphs, or Opt-in Slicing Mode.

1. Capture raw intent.
2. Architect AI writes mission brief and task graph.
3. Human owner approves the slice or task package.
4. Worker agents run on separate branches or worktrees.
5. Each worker writes a run log.
6. QA Agent reviews validation and visual checks.
7. Integrator Agent reads diffs and recommends merge order.
8. Human owner decides merge and release.
9. Release plan and roadmap status update.

Parallel Work Rules

Parallel workers must have disjoint allowed paths.
Shared contract work must run before dependent UI/API/app work.
Integrator Agent reviews all branches before merge.
If two workers touch the same file, stop and re-plan.
If a worker needs forbidden paths, it must report scope expansion instead of editing.

Branch And Run Discipline

Each worker task should record:

Branch or worktree.
Source prompt ID.
Task graph link.
Allowed paths.
Forbidden paths.
Changed files.
Validation commands.
Risks and follow-up.

Warp-era Terminal Workflow

Warp, terminal agents, or local command agents can act as a cockpit for:

Running pnpm quality commands.
Inspecting logs.
Creating previews.
Checking local build artifacts.
Running safe scripts and runbooks.

Codex, Cursor, Claude Code, and similar tools are coding workers. CI, Storybook, UI Workshop, previews, and quality checks are the feedback layer.

Ops Agent must not manage secrets, write tokens into the repo, or mutate production infrastructure without explicit human approval.