| # Benchmark Optimization Loop |
|
|
| ## Objective |
|
|
| Improve a measurable system outcome through small experiments while preserving correctness and keeping every accepted change reproducible. |
|
|
| ## Use This When |
|
|
| - You have a stable benchmark, eval suite, or objective metric. |
| - One change can be tested independently of the next. |
| - The system can reject regressions and restore the last accepted state. |
|
|
| Do not use this loop when the score is easy to game, the benchmark is still changing, or the real outcome requires human judgment that the metric does not capture. |
|
|
| ## Trigger |
|
|
| - Schedule: a bounded overnight or weekly experiment window. |
| - Event: a new baseline, model, dataset, or optimization target is available. |
| - Manual bootstrap: "run up to five benchmark-backed optimization experiments." |
|
|
| ## Intake |
|
|
| - Baseline artifact, benchmark command, correctness checks, and target metric. |
| - Prior experiment ledger, failed hypotheses, and protected files. |
| - Compute, token, time, and concurrency budgets. |
|
|
| ## Agents |
|
|
| - Experiment designer: proposes one falsifiable change and expected effect. |
| - Implementer: applies only that change in an isolated candidate workspace. |
| - Verifier: runs correctness checks and the frozen benchmark. |
| - Recorder: accepts or rejects the candidate and updates the experiment ledger. |
|
|
| ## Workspace And Permissions |
|
|
| - Use a disposable worktree, branch, or sandbox per candidate. |
| - Allow edits only to the declared optimization surface. |
| - Keep benchmark data, scoring code, holdout cases, and acceptance thresholds read-only. |
| - Disallow test deletion, scorer edits, hidden-test inspection, and concurrent candidates that exceed the budget. |
|
|
| ## Durable State |
|
|
| - Baseline version and score, hypothesis, candidate diff, commands, raw results, cost, decision, and rejection reason. |
| - The last accepted artifact remains the parent of the next experiment. |
|
|
| ## Loop Steps |
|
|
| 1. Freeze the baseline, benchmark version, correctness gate, and budget. |
| 1. Read the ledger so the next hypothesis does not repeat a failed experiment. |
| 1. Propose one bounded change with a predicted effect. |
| 1. Apply it in an isolated candidate workspace. |
| 1. Run correctness checks before the benchmark. |
| 1. Compare repeated benchmark runs against the accepted baseline. |
| 1. Accept only a reproducible improvement that clears the minimum delta; otherwise reject and restore the baseline. |
| 1. Record the full receipt and continue until the target or budget is reached. |
|
|
| ## Verification Gates |
|
|
| - Correctness and safety checks pass unchanged. |
| - The benchmark, dataset split, scorer, and environment match the recorded baseline. |
| - Improvement clears the declared minimum delta across the required repeats. |
| - The candidate does not worsen protected secondary metrics beyond tolerance. |
| - The ledger contains enough evidence to reproduce the decision. |
|
|
| ## Budget And Exit |
|
|
| - Max retries: 5 candidate experiments. |
| - Max runtime: 240 minutes. |
| - Stop on target attainment, budget exhaustion, two repeated hypotheses, benchmark instability, or a protected-metric regression. |
|
|
| ## Escalation |
|
|
| Escalate when the metric conflicts with observed quality, a candidate changes the evaluator, results vary beyond tolerance, or the next experiment would broaden permissions or compute. |
|
|
| ## Loop Instruction |
|
|
| ```text |
| Optimize <artifact> against <benchmark> for at most five experiments. |
| Treat <correctness command> as a non-negotiable gate and <metric> as the optimization signal. |
| Change one declared variable per candidate in an isolated workspace. Never edit the |
| benchmark, scorer, holdout set, or protected tests. Accept a candidate only when repeated |
| runs improve the baseline by <minimum delta> without regressing <secondary metrics>. |
| Record every hypothesis, diff, command, score, cost, and accept/reject decision. |
| ``` |
|
|
| ## Worked Example |
|
|
| A team wants to reduce an agent workflow's median latency without lowering task success. The loop starts from a frozen 78% success / 42-second baseline, tries one routing or caching change at a time, reruns the same 100-task suite three times, and accepts only candidates that keep success within one percentage point while cutting median latency by at least 5%. |
|
|
| ## Failure Modes |
|
|
| - Optimizing on the same cases used to invent the change. |
| - Accepting a noisy single run as improvement. |
| - Changing multiple variables and losing causal attribution. |
| - Improving the headline metric while silently degrading cost, safety, or tail latency. |
| - Letting the agent modify the verifier that judges its own work. |
|
|
| ## Example Contract |
|
|
| - [`examples/benchmark-optimization-loop.json`](../examples/benchmark-optimization-loop.json) |
|
|
| ## References |
|
|
| - [A Self-Improving Coding Agent](https://arxiv.org/abs/2504.15228) - Demonstrates benchmark-gated self-modification with measurable gains. |
| - [Understanding the Challenges in Iterative Generative Optimization with LLMs](https://arxiv.org/abs/2603.23994) - Identifies evaluation and credit-assignment choices that make iterative optimization brittle. |
|
|
