Define Agent Team

Expected Outputs

• updates to agents/ when the shared role layer changes
• updated repo-scoped Claude agent definitions when supported
• documented Codex local config guidance when .codex is intentionally local-only
• documented Cursor custom-mode and background-agent guidance when manual UI setup is still required
• explicit guidance on when to use a full parallel team versus an advisor loop
• a concise explanation of how to launch or use the resulting team

Default Pattern

1. Read the existing role briefs in agents/.
2. Verify the current official docs for each target client.
3. Keep agents/ as the conceptual source of truth for role purpose and output shape.
4. Add or update repo-scoped client definitions where the platform supports them.
5. Document manual steps explicitly where the platform still expects UI configuration.
6. Decide whether the repo should prefer a parallel specialist team, an advisor loop, or both.
7. Keep model defaults explicit when the platform supports stable names.
8. Keep file ownership separate so parallel workers do not collide.

Validation Checklist

• Confirm the role system stays conceptually aligned across Claude Code, Codex, and Cursor.
• Confirm repo-tracked files only include configurations the target client actually supports.
• Confirm manual setup steps are documented where automation is not supported.
• Confirm the recommended team shape is the smallest coherent one for the repo.
• Confirm model defaults are explicit where possible and not brittle where the platform is dynamic.
• Confirm ownership and handoff guidance are clear enough to support parallel work safely.

Guardrails

• Do not invent unsupported config formats.
• Do not let Claude, Codex, and Cursor baselines drift into three different role systems.
• Do not hardcode brittle Cursor model names when the official docs treat model availability as dynamic.
• Do not replace a local-only .codex convention with tracked repo files unless the repo explicitly chooses to do that.
• Do not use a full team when one executor plus one advisor would be simpler and safer.
• Do not optimize for maximum agent count; prefer the smallest coherent team.
• Do not hide manual setup steps when a client still requires them.