2.5 Evaluating the Impact of Reflection

📄← Module index↗ · 📄↑ Master index↗

After this section, you should be able to:

• Explain what evaluating the impact of reflection contributes to an agentic system.
• Decide when to use it and when a simpler design is sufficient.
• Implement the pattern as observable, testable components.
• Identify its principal cost, safety, and reliability risks.

Evaluate reflection when deciding whether an extra critique-revision cycle belongs in production.

The correct question is not whether the revision sounds better, but whether it improves defined metrics reliably.

The arrows show control and data movement, not necessarily separate models. A deterministic function, one model called multiple times, or several models may implement the boxes. Preserve trace identifiers across the flow.

1. Create examples with known failure modes. Write the input, expected output, owner, and failure condition before implementation.
2. Run direct and reflected variants. Keep the decision observable in logs or structured state so it can be evaluated.
3. Score both with human or deterministic checks. Apply least privilege and validate assumptions at this boundary.
4. Segment results by error type. Capture the result and enough metadata to reproduce or diagnose it.
5. Review regressions introduced by revision. Compare the result with explicit acceptance criteria before continuing.
6. Set a maximum reflection budget. Route failures to retry, fallback, or human review according to policy.

• State: Keep messages, tool calls, observations, artifacts, decision reasons, attempt count, token use, latency, and final status in a structured run record.
• Contracts: Define each component with typed inputs, typed outputs, allowed side effects, timeouts, and error categories.
• Controls: Use least-privilege credentials, allowlisted tools, bounded loops, input validation, output validation, and approval gates for consequential actions.
• Observability: Record prompts or prompt versions, model and parameters, tool arguments, tool results, exceptions, timestamps, and cost—subject to privacy rules.
• Evaluation: Test representative, edge, adversarial, and failure-recovery cases. Compare against the simplest viable baseline.

Track factual accuracy, instruction compliance, style, latency, tokens, and revision-induced errors.

INPUT: user goal + constraints
STATE: {run_id, step, observations, budget, status}
DECIDE: next bounded action
VALIDATE: permissions, arguments, and policy
EXECUTE: model call, deterministic code, or tool
OBSERVE: structured result or categorized error
STOP: acceptance criteria pass, budget reached, or human escalation

• The use case justifies this pattern over a simpler one-shot call.
• Inputs, outputs, and success criteria are explicit.
• Tool calls and side effects are validated and permissioned.
• Loops have hard budgets and meaningful stop conditions.
• Failures produce safe retries, fallbacks, or escalation.
• Quality, latency, and cost are measured together.
• Regression tests include at least one failure case.

1. What observable failure would show that this pattern is misapplied?
2. Which part should be deterministic rather than delegated to an LLM?
3. What is the minimum context required at each step?
4. Where should a human approval or escalation gate sit?
5. Which metric would prove that the added complexity is worthwhile?

Implement a minimal version for one narrow task. Save three traces: a successful run, a recoverable failure, and a case that must stop or escalate. Write one regression test for each trace and compare the result with a direct-generation baseline.