Trainer Train - Agent

Use this skill as the orchestration contract for one trainer run against an agent contract target: a *.agent.md file, custom agent definition, or agent instruction document that configures tool routing, MCP skill invocation, agent personas, or handoff behavior.

Read references/agent-loop-contract.md for the full routing table, agent-contract constraints, tool-routing rules, and validation requirements before any stage execution.

When to use this skill

• The selected target is an agent contract file (*.agent.md) or agent instruction document.
• The caller needs to initialize or resume a trainer workspace for an agent target.
• Known failure modes include wrong tool routing, over-broad persona, poor MCP skill invocation, or unbounded handoffs.
• Missing datasets or eval manifests need to be synthesized before optimization.
• A winning candidate needs to be validated and written back to the source agent contract.

Do not use this skill for pure prompt files, Python code targets, or skill files. Read the parent trainer skill's references/target-routing.md to identify the appropriate specialist for those target types.

Required inputs