Skip to content

Convergence Detection in Iterative Agent Refinement

Monitor three observable signals across refinement passes to replace intuition-based stopping with a mechanical criterion.

The Problem

Iterative refinement loops — plan polishing, critique passes, bead polishing, documentation drafts — have no natural stopping point. Agents and developers either stop too early (leaving unresolved issues) or over-refine (wasting compute on passes that change nothing). "It looks good enough" is not a stopping criterion.

For tasks with a test harness, this is solved: tests pass → stop — the PASS/FAIL gate an evaluator-optimizer loop leans on. For prose, specs, and design documents, no such machine-checkable gate exists. Convergence detection fills that gap.

The Three Signals

Monitor these signals across consecutive refinement passes:

Signal Converging Diverging
Change velocity Rate of modifications slows — pass N changes 30%, pass N+3 changes 2% Rate stays high or accelerates
Output size Size stabilises or shrinks — additive passes are exhausted Size grows — indicates scope creep, not refinement
Content similarity Diff between consecutive passes shrinks toward zero Diff stays large — substantive issues remain unresolved

When all three signals converge simultaneously, further passes yield diminishing returns. When any signal diverges, issues remain unresolved and more passes are warranted.

Failure Patterns

Three patterns indicate a restart is needed rather than continued iteration:

  • Oscillation — output alternates between two versions across passes; the agent cannot resolve a trade-off without external input
  • Expansion — output grows each pass instead of shrinking; scope is drifting rather than stabilising
  • Low-quality plateau — all three signals converge but output quality remains poor; the approach needs redesign, not more passes

Five-Pass Blunder Hunt

For critical outputs — major design specs, agent system prompts, architectural decisions — run the identical critique prompt five consecutive times against the same output. Each pass surfaces issues that previous passes normalised over. A single critique pass produces false confidence; repeated identical passes force examination of progressively subtler problems.

This technique applies the convergence signals: if pass 4 and pass 5 produce nearly identical critiques with no new issues, content similarity has converged and the output is stable.

Relationship to Other Stopping Mechanisms

Mechanism When to use
Convergence detection Prose, specs, design docs — no test harness available
PASS/FAIL from evaluator Code tasks with executable tests — machine-checkable
Max round limit Fallback for all loops — prevents runaway iteration
Model self-declaration Low-cost tasks where precision matters less

Convergence detection complements the evaluator-optimizer pattern's max-round fallback: the evaluator-optimizer terminates on PASS or round limit; convergence detection tells you when to set that round limit or when to stop early without a formal evaluator.

Production tools increasingly pair an evaluator with a hard round cap rather than relying on either alone. Microsoft's VS Code ships an Advanced Autopilot mode whose utility-model judge decides loop completion by reading the run transcript, bounded by a maximum of three loops (VS Code 1.124 release notes). This couples a transcript-aware, evaluator-style stopping decision with the max-round fallback.

Example

A developer is running critique passes on a system prompt for a coding agent. After each pass they compare the new version against the previous.

Pass 1 → Pass 2: 40% of lines changed. Output grew by 200 words. Clear convergence signal: diverging.

Pass 3 → Pass 4: 15% of lines changed. Output size stable. Partial convergence.

Pass 4 → Pass 5: 3% of lines changed (minor phrasing only). Output size unchanged. Diff near-zero. All three signals converge: stop.

Running a sixth pass — one beyond the five-pass blunder hunt — would likely produce cosmetic changes that may degrade quality by introducing unnecessary variation.

When Signal Convergence Misleads

Convergence signals measure whether the output is stabilising, not whether it is correct. Lee et al., RefineBench: Evaluating Refinement Capability of Language Models via Checklists (2025), evaluated Gemini 2.5 Pro, GPT-5, and DeepSeek-R1 on 1,000 problems across 11 domains and found that self-refinement without external feedback yielded gains of +1.8 percentage points or less over five iterations, while guided refinement approached near-perfect scores — and that models routinely halt early due to overconfidence even when errors remain. When signals converge without an external evaluator, the stable state may reflect self-bias, not quality. For high-stakes outputs, pair convergence detection with an external checker (tests, a second model, a human reviewer) rather than relying on the signals alone.

Key Takeaways

  • Three signals — change velocity, output size, content similarity — replace intuitive stopping with observable criteria
  • Oscillation, expansion, and low-quality plateau are failure patterns that require a restart, not more passes
  • The five-pass blunder hunt applies convergence detection to critique loops: when consecutive passes produce near-identical critiques, the output has stabilised
  • Convergence detection fills the gap for prose and design tasks where no test harness exists; use PASS/FAIL from tests for code
  • Always pair with a hard max-round limit as a cost fallback

Sources

Adopted — practitioner-sourced, corroborated Last reviewed: 2026-06-14
Feedback