A reviewed multi-agent workflow for Claude Code and Codex
Plan the dependencies, file ownership, checks, and workspaces before dispatch. Start only work whose dependencies are ready, integrate it explicitly, and have someone else review the combined result.
Choose several agents for dependency structure, not task size.
A large task is not automatically a multi-agent task. If several writers need the same files or if every decision depends on the previous one, a single well-steered session may be faster and safer. Multiple agents help when useful work can proceed independently and their outputs can be integrated through explicit contracts.
| Mode | Best fit | Primary risk |
|---|---|---|
| One session | Sequential decisions, one tightly coupled change, or a task needing one coherent context. | A long thread can become hard to inspect or recover. |
| Parallel sessions | A few independent tasks you already understand and can supervise directly. | Overlapping writes without real directory isolation. |
| Dependency graph | A larger objective needing research, synthesis, implementation, integration, and cross-review. | An unreviewed plan can scale a bad assumption across every task. |
Parallel agents are not automatically isolated. Concurrent writers need separate real directories—usually worktrees—unless their dependencies serialize the writes.
Review the plan as an execution contract.
The useful planning artifact is not a paragraph of encouragement. It is an editable graph that names what every task receives, changes, checks, and publishes.
- Provider and roleChoose Claude Code or Codex deliberately. Research and review roles need a runtime that can enforce read-only behavior.
- Prompt and acceptance checksState the bounded objective and the evidence that proves completion.
- Dependencies and artifactsName what must exist before a task starts and what versioned output it publishes for consumers.
- File claims and workspaceExpose expected write areas and the actual directory or detached worktree before two agents can collide.
- Integration and cross-reviewReserve explicit tasks to assemble results and independently challenge the integrated outcome.
Treat approval as permission for the reviewed dispatch contract, not blanket approval for whatever an agent later decides to do. A breaking artifact change should pause consumers until they acknowledge the new version.
Dispatch readiness, not everything at once.
Plan read-only
A lead turns the objective into proposed research, synthesis, implementation, integration, and review tasks without changing the project.
Approve or edit the graph
The operator adjusts prompts, providers, dependencies, checks, file claims, and worktree paths before execution.
Start dependency-ready tasks
Independent tasks can proceed together. Consumers wait until required artifacts exist and accepted versions are explicit.
Steer individual work
Stop, retry, reroute, or add guidance to one task without discarding unrelated completed work.
Integrate deliberately
A named integration task combines outputs and runs system-level checks. Individual task success is not product success.
Cross-review independently
A separate reviewer examines the integrated result and its evidence. Only completed integration and cross-review can allow Done.
Do not convert uncertainty into a green badge.
A task can fail its checks, lose its runtime, publish an incompatible artifact, or return a dirty worktree. Each state needs an operator-visible consequence. Retry when the contract remains valid; reroute when another provider is a better fit; revise the graph when the dependency itself was wrong.
- Overlapping writes: warn before launch; serialize dependencies or provide separate workspaces.
- Dirty worktrees: hold them for review. Crowsnest never removes, resets, or cleans an agent worktree.
- Interrupted runs: persist the execution state and return unfinished work for human review after restart.
- Missing review: remain incomplete. “Agents finished” is not a substitute for integration and cross-review.
OpenAI describes Codex threads as resumable and forkable units in its app-server architecture. Keep provider-native histories separate and carry attribution in review artifacts rather than pretending several agents share one merged model thread.
Crowsnest retired its former Crew orchestrator in October 2026. It now supervises ordinary provider-native conversations, jobs, and subagents instead of dispatching this graph itself.
Sources and related field notes
Workflow references: OpenAI's Codex app overview, Codex app-server architecture, and Conductor's parallel-agent concepts. The current Crowsnest boundary was checked against the released source on October 19, 2026.
Published August 24, 2026; current-product claims reviewed October 19, 2026. Provider runtimes evolve independently; linked first-party documentation is the current source for their native capabilities.