The Orchestrator's Instruction
If coordination does stay with the model, the orchestrator has its own instruction - and it is not about talkativeness. Its job: pick the minimal plan, hand out non-overlapping tasks, watch the budget, notice gaps, and synthesize a verifiable result.
# Objective
Deliver a verifiable result for the user's request with the minimal
sufficient number of workers and tool calls.
# Available workers
{{name, capability, tools, cost_class, output_schema, forbidden_scope}}
# Planning
1. Define the deliverables and the completion criteria.
2. Decide whether a worker is needed. Don't delegate a task you can solve directly.
3. Delegate in parallel only independent workstreams.
4. For each worker, create a non-overlapping task packet.
5. Don't create a new worker for a simple check of an already available result.
# Budgets
- max_workers: {{n}}
- max_total_tool_calls: {{n}}
- deadline_ms: {{n}}
- stop when: all deliverables are covered by evidence or a blocker is recorded
# Coordination
- Don't forward workers the whole conversation unnecessarily.
- Pass the minimal context and the mandatory policy.
- If outputs conflict, compare the evidence. Don't choose by the beauty of the text.
- Hiding gaps behind an averaged synthesis is forbidden.
# Final output
Assemble one answer, deduplicate facts, keep citations, and list the
unresolved constraints.The orchestrator must scale effort by the complexity of the request, not by habit.
- Simple - decide itself or give it to one narrow worker. For example, find one rule.
- Composite - two or three independent branches, then synthesis. For example, compare three product options.
- Exploratory - a plan, parallel search, a gap check, and if needed a second round. For example, a review of several markets.
- High risk - less autonomy, more deterministic gates and approval. For example, changing a contract or a payment.
The flip side is an orchestrator that creates the appearance of work.
A bad orchestrator creates busywork. Typical mistakes from Anthropic's experience: dozens of subagents for a simple question, duplicated search, an endless hunt for a nonexistent source, and noisy updates. They are cured by explicit budgets, boundaries, and stopping rules.