Reflection without explicit stopping criteria is recursive cost, not self-improvement

Reflection without explicit stopping criteria is recursive cost, not self-improvement. Asking an agent to critique, revise, and try again creates another inference loop; it does not create a learning system by itself. Unless the loop can distinguish progress from paraphrase and knows when further work is worth its price, “reflect more” is just an unbounded token budget wearing the language of metacognition.

The reflective-agent literature contains a real mechanism. Reflexion stores verbal feedback in episodic memory, Self-Refine iterates through critique and revision, and credit assignment attempts to connect an outcome to the decisions that produced it. These patterns can improve a later attempt without changing model weights. But each one depends on a signal that the next attempt is better. If the same model proposes, judges, and rewrites against a vague instruction, confidence can rise while quality stands still.

I want three stopping conditions before I enable reflection. First, a hard resource boundary: maximum iterations, tokens, elapsed time, and tool spend. Second, a quality boundary: executable tests, a rubric with anchored thresholds, or a verifier independent enough to reject cosmetic change. Third, an escalation boundary: conditions under which the system admits that more autonomous work is unlikely to help and hands the task to a person or a different capability. These are not emergency brakes. They define the algorithm.

Marginal improvement should be observable. Preserve each candidate, critique, score, and action so the loop can compare deltas rather than trust its latest prose. Stop when tests pass, when the score clears the task’s threshold, or when improvement over two rounds falls below a defined amount. Stop sooner when revisions oscillate between alternatives, repeat the same failure, or consume a growing budget without changing externally visible behaviour. That turns “self-improvement” into a bounded search process.

Supervision level should change with reversibility. Let an agent reflect freely while drafting, simulating, or exploring a sandbox. Require a human decision before an irreversible effect even if the reflective loop reports confidence. The Vidhyarthi supervision material makes reversibility the useful boundary: access, time, and authority must make oversight meaningful, not ceremonial. A final self-critique cannot substitute for approval when the action sends money, deletes data, or publishes externally.

This also changes how I account for cost. Cost is an architectural property, and reflection multiplies the most expensive component by design. The budget should therefore be allocated by failure value, not enabled globally. A second pass may be cheap insurance on a consequential contract analysis and waste on a reversible classification. Extended thinking, episodic memory, and extra reviewers all compete for the same latency and spend envelope.

There is one precise concession: open-ended reflection is useful during research or deliberate exploration when discovering the evaluation criteria is itself part of the task. In that setting, novelty and reframing can justify a wider loop. It still needs a session budget and a human who decides whether the exploration has produced something worth keeping. Uncertainty about the destination is not permission for infinite travel.

I do not reject reflection; I reject treating recursion as evidence of learning. A production system should be able to show why another iteration was authorized, what changed, how improvement was judged, and why the loop stopped. Without those answers, the agent has not become more capable. It has become more expensive and harder to supervise.