Customize an agent

Agents are markdown procedures sourced from agents/<name>.md. The Planner, Coder, Evaluator, and every other workflow role simply load the resolved file paths from _intelligence — they do not hard-code behaviour beyond tool wiring (/concepts/runner-spec/).

Where overrides live

1. Tenant overlay — agents/ in your team intelligence repo (/guides/tenant-overlay/).
2. Repo overlay — .coro/agents/ in the target checkout (/guides/repo-overlay/).

Relative paths (agents/foo.md) are merged with replace semantics (last-wins from the resolver stack):

base (@coro-ai/intelligence-base/layer/agents/foo.md)
  → tenant/agents/foo.md (if present)
    → repo .coro/agents/foo.md (if present)

If only the tenant supplies agents/foo.md, it fully replaces the base file for that path (not piecemeal section merge).

Note: .claude/CLAUDE.md is not mirrored here — CLAUDE.md merges by append, not wholesale replace (/concepts/architecture/).

Typical customisation axes

Teams usually touch:

Area	Reason
Extra guard language	Codify SOC2 / HIPAA language for review phases.
Tool usage policy	Narrow when planners may skip skills or escalate.
Diff / PR etiquette	Tie into your branching model or changelog rules.

Keep overrides delta-sized: copy the upstream base file once, then minimise drift so rebasing intelligence across Coro upgrades stays mechanical.

How propose_change fits

Agents should not rewrite intelligence from the sandbox. Approved updates flow through propose_change (opens a Git PR). Reviewers gate changes in the tenant or repo overlay Git remote — identical human workflow to application code (/guides/review-proposals/).

Related reading

• /guides/add-skill/ — prefer skills for encyclopaedic conventions; reserve agent edits for flow control.
• /guides/switch-lanes/ — specialised agents differ per FAST / STANDARD / DEEP workflow.