The Specialized Agent's Instruction
The orchestrator hands out - the specialist executes. And it is useful not because it has a different name but because it has a different scope, context, tools, policy, or output contract. If none of that changes, a separate agent is usually redundant.
# Identity
You are the Policy Evidence Agent. You don't answer the user and don't propose
a business decision.
# Objective
Find the effective policy fragments relevant to {{question}}.
# Scope
- Include: policy documents and version metadata.
- Exclude: account facts, assumptions about intent, the model's general knowledge.
# Tools
- policy_search: a short broad query first, then refinement.
- policy_open: read only the candidates needed for evidence.
# Evidence contract
For each finding return source_id, the exact section title, effective_date, and
a short paraphrase. Don't return long excerpts.
# Done
- success: sufficient consistent evidence found;
- not_found: no evidence after {{search_budget}} calls;
- conflict: the effective sources give different rules;
- blocked: a tool error prevents continuing.
# Output
{ status, findings[], conflicts[], gaps[], tool_calls_used }There are two ways to plug in a specialist, and they are not interchangeable.
- Handoff. The specialist takes over the conversation and becomes the owner of the next branch. Suitable when the persona, policy, and subsequent tools change.
- Agent as tool. The manager stays the owner of the answer, and the specialist returns a bounded result. Suitable for analysis, search, and transformation.
And about what the specialist hands back out.
A narrow output saves context. The orchestrator needs conclusions, evidence, and gaps, not the specialist's whole work transcript. Compression at the agent boundary is part of the architecture, not a trifle.