docs / concepts / plan-mode

Plan mode

Multi-worker orchestration for big tasks. The Conductor plans, Workers execute in parallel, and a synthesis step combines the results into one answer with provenance.

When you want it

Use plan mode when the task is too big for a single model in a single pass: "audit this codebase for security issues", "design and scaffold a new feature end-to-end", "review this architecture and propose alternatives". The Conductor sees the whole goal, the Workers see only their slice.

The pipeline

  1. Conductor receives the goal, generates a plan as a directed acyclic graph of tasks. Each task has a role, a goal, dependencies and a tool allowlist.
  2. Workers pick up their tasks. Each Worker is an agent persona matched to the task role (backend-developer, code-reviewer, devops-engineer, etc.). Tasks with no dependencies run in parallel.
  3. Synthesis combines the WorkerResults. Strategies: merge (concatenate with section headers), consensus (vote on contested points), select_best (pick the strongest answer).
  4. Final answer comes back with full provenance: which Worker did what, which model each one used, and how long each step took.

Worker roles

Workers are typed by role. Each role has a capability profile, a tool allowlist and a default system prompt.

Role definitions live in config/worker_roles.yaml. You can add or override roles there.

Cost balance

Plan mode runs multiple models, which costs more than a single completion. The Conductor estimates total cost before dispatch and can be capped per request. You can also pin plan mode to local_only, in which case it only uses Ollama models and costs nothing in API fees.

Use it sparingly
Plan mode is powerful but not free. Use it for security reviews, architecture decisions, and anything else where the cost of being wrong is higher than the cost of running three models in parallel. For everything else, single-model routing is faster and cheaper.

API reference: worker endpoint →