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
