liveBook · Manning

Reflection

As explained in Designing AI Agents

An agent capability for inspecting and revising its own intermediate results or proposed actions.

Reflection connects feedback to a permitted change in an agent’s work. It can revise a current artifact or execution, or influence later tasks through an admitted capability or lesson. Design it around three questions: what evidence can the feedback see, when will the result become known, and what may the agent change? section 7.1 section 7.9

Reflection adds a feedback path to perception, memory, reasoning, and action. After an action, it can compare the result with the intent, return work to reasoning, revise the artifact, or stop. In a controlled rollout, for example, reflection checks the result against both the requested feature and the project rule; collaboration can provide another point of view. section 2.2.1 section 2.1.1

Four patterns differ in what they can change and how long the change persists:

  • Generator-Critic reviews and revises the current artifact against criteria, source material, tests, or tool results, then accepts, stops, or exhausts its review budget. section 7.1
  • Self-Heal Loop repairs current execution and bounded working state. Verify each repair, preserve rollback, and hand off when the recovery limit is reached. section 7.1 section 7.9
  • Skill Package admits a reusable capability for later tasks after isolated, coexistence, and trial evidence. Routing filters by permission, risk, dependency, and applicability. section 7.1 section 7.9
  • Experience Replay separates raw traces, candidate lessons, and admitted advice. Recording, admission, and runtime recall are distinct decisions; recalled experience remains advisory. section 7.9
Table 7.2 Evidence and authority across the four patterns
Pattern
Primary evidence
Change authority
Persistence
Main stop
Generator-Critic
Criteria, source material, tests, or tool results
Revise the current artifact
Current delivery
Accept, stop, or exhaust review budget
Self-Heal Loop
Failure signal, protected verifier, and regression results
Repair bounded execution and working state
Current task
Recover, roll back, or hand off
Skill Package
Isolated, coexistence, and trial evidence
Admit and route a versioned capability
Across tasks
Activate, reject, deprecate, or retire
Experience Replay
Traces, independent cases, causal evidence, and later outcomes
Admit and recall scoped advice
Across tasks within scope
Replay, withhold, conflict, or retire

Compare each pattern’s evidence, change authority, persistence, and stopping decisions. section 7.1

Another model pass is not verification. Without new evidence, reconsideration can improve an answer or turn a correct answer into an incorrect one. A useful critic has a different review frame, criterion, context, or evidence—not merely a different model identity. Tests, tool results, source documents, human judgment, and observed outcomes can ground review, but each establishes only claims within its scope. section 7.1.1 section 7.1

When local repairs accumulate

An environment-building agent added a safeguard after each build failure. Each fix addressed a local problem, but after roughly two months the accumulated checks made builds rise from about five minutes to thirty minutes, sometimes nearly an hour. Cross-run review is needed to measure combined cost, merge overlapping rules, and remove rules that no longer help. section 7.0

Evaluate changes at the scope they affect. A passing unit test does not establish that accumulated repairs improve the whole system. Track regressions, cost, latency, state changes, and human effort alongside task success; do not let an aggregate score hide a hard regression. section 7.1 section 7.9

How Designing AI Agents covers it

How deep it goes
Develops four patterns through evidence, change authority, persistence, stopping decisions, and evaluation signals.
Examples
Uses a controlled rollout and an environment-building agent whose accumulated safeguards slowed builds.
What you can do after
Evaluate and understand feedback paths by matching evidence to change authority, stopping conditions, and cross-run effects.
How it connects
Connects reasoning, action, memory, collaboration, governance, and evaluation through feedback and controlled changes.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.