---
title: "Coaching / Meta-Supervisor · agent topology · Agent Fieldbook"
url: https://agentfieldbook.org/design/topologies/coaching/
description: "A read-only meta-agent sits above the main agent, observing actions and providing advisory feedback. Main agent decides whether to act."
section: "Design · Topologies"
source: Agent Fieldbook — generated from the published page
---

# 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](https://aclanthology.org/2026.acl-demo.34/)

- [aclanthology.org](https://aclanthology.org/2026.acl-demo.34/)

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.

## In the anatomy

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.

## At a glance

| Field | Value |
| --- | --- |
| Pattern | Coaching / Meta-Supervisor |
| State model | session |
| Example | LangGraph human-in-the-loop patterns· Claude teach mode |

## Common use cases

## Strengths

## Weaknesses

## Basis

Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.

## Verified against

- [https://aclanthology.org/2026.acl-demo.34/](https://aclanthology.org/2026.acl-demo.34/)

The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.