Coaching / Meta-Supervisor
A read-only meta-agent sits above the main agent, observing actions and providing advisory feedback. Main agent decides whether to act.
Verified against aclanthology.org
A coaching topology places a read-only meta-agent above the working agent: it observes actions and outcomes and feeds back improvements, better prompts, corrected steps, updated policy, without taking over execution. It separates doing from improving, so the working agent stays simple while a supervisor drives learning over time. In anatomy terms it is a second core with a view of the first's memory and trace, closing a slow outer loop.
A meta reasoning core observes the working agent's trace and memory read-only, then writes back guidance or policy on a slow outer loop; it does not sit in the fast execution path; the perimeter governs any policy it changes.
- ·Continuous-improvement loops: reflection and optimiser layers.
- ·Prompt and policy tuning from observed runs.
- ·Quality supervision of a fleet of working agents.
- ·Separates execution from improvement, so each stays simple.
- ·Drives learning without destabilising the live agent.
- ·One supervisor can coach many workers.
- ·Safer than self-modification in the hot path.
- ·Slow outer loop: improvement lags behind the runs.
- ·The coach's judgment can itself be wrong.
- ·Extra cost for the supervisory pass.
- ·Feedback must be applied carefully to avoid drift.
Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.
The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.