Artifact registry
Immutable candidates, parent links, diffs, prompts, models, datasets, traces, scores, and environment hashes.
Research notes · August 2026
Building bounded self-improving software loops.
A practical architecture derived from the LoopClimber tour and grounded in program search, software-agent evaluation, metric-driven pipeline optimization, evaluator calibration, and AI risk management.
Core principle
“Self-improving” should mean searching over versioned, inspectable software artifacts under fixed tests and explicit approval boundaries—not unconstrained recursive self-modification.
01 / Operating model
A general loop, from baseline to guarded promotion.
The page’s candidate sub-DAG, eval, cluster, pruning, and regression-bank ideas map cleanly to an evolutionary search system. The critical addition is a hard boundary between the search process and the production promotion process.
-
01
Define the improvement contract.
Write down the task distribution, mutable surface, non-negotiable invariants, resource budget, stop conditions, and who may approve a release. Change prompts, routing, tools, models, or code only inside that envelope.
-
02
Freeze a reproducible baseline.
Version the workflow, environment, dependencies, datasets, model identifiers, seeds, and traces. Separate search data from validation and held-out promotion cases so the loop cannot merely memorize its evaluator.
-
03
Generate bounded, diverse candidates.
Propose small diffs against the baseline and preserve provenance: parent, mutation, rationale, model, prompt, and cost. Sample from more than the current winner to avoid collapsing into a local optimum.
-
04
Run candidates in isolation.
Execute matched cases in sandboxes with least-privilege tools, network and filesystem boundaries, timeouts, and cost limits. Parallel workers increase throughput; identical inputs make comparisons meaningful.
-
05
Evaluate through layered gates.
Start with executable checks—build, types, unit and property tests, invariants, security, and policy. Then measure quality, robustness, latency, and cost over repeated trials. Use calibrated model judges or humans only where code cannot decide.
-
06
Keep a diverse archive.
Compare candidates on a Pareto frontier rather than one blended score. Retain useful specialists and stepping stones across quality, speed, cost, risk, and behavioral niches; prune only dominated or invalid branches.
-
07
Expand the regression bank.
Turn every newly observed failure into the smallest reproducible case and tag the affected slice. Keep regression tests separate from the held-out promotion set, and audit for leakage and redundant cases.
-
08
Require governed promotion.
Promote only after held-out gates, uncertainty checks, human approval, and a reviewable diff. Release gradually with canaries, monitoring, a kill switch, and automatic rollback. The winner becomes the next baseline; the safety envelope does not mutate itself.
02 / System design
The smallest useful architecture.
Proposal engine
Model or human generators constrained to declared mutation operators and the editable surface.
Isolated runners
Parallel, reproducible sandboxes with bounded tools, compute, data access, wall time, and spend.
Evaluator DAG
Cheap deterministic gates first; slower statistical, qualitative, adversarial, and human checks later.
Archive + selector
Pareto ranking, diversity preservation, uncertainty handling, and explicit reasons for pruning.
Promotion controller
Held-out evaluation, approval, signed release, staged rollout, telemetry, rollback, and a full audit trail.
Keep the evaluator outside the candidate’s write boundary.
A candidate must not alter its tests, held-out cases, score aggregation, audit log, permissions, or promotion rules. Changes to that control plane belong to a separate, human-governed process.
03 / Measurement
Use a promotion contract, not a leaderboard vibe.
Most software changes are multi-objective and many AI measures are noisy. Record distributions and confidence bounds, compare on matched cases, and state trade-offs directly.
- Correctness
- Pass rate, invariant violations, regression slices, task completion
- Robustness
- Adversarial cases, perturbations, tool failures, recovery behavior
- Quality
- Human rubric, calibrated judge agreement, citation or evidence checks
- Efficiency
- p50/p95 latency, tokens, tool calls, compute, dollars, energy proxy
- Safety
- Unsafe actions, boundary attempts, sensitive-data exposure, rollback readiness
- Search health
- Diversity, improvement per trial, evaluator failures, archive coverage
Example decision rule
promote = hard_gates_pass
AND quality_delta_lower_bound > 0
AND p95_latency <= budget
AND cost <= budget
AND safety_events == 0
AND human_approval == true
Tolerances are domain-specific. Predeclare them before generating candidates, and treat a judge score as a measurement with error—not ground truth.
04 / Research map
Search terms from the tour.
These queries translate the page’s visual vocabulary into established research language and are a useful starting point for deeper literature searches.
"LLM workflow optimization" evaluator computational graphexecutable specification property-based testing AI agentsasynchronous evolutionary program search parallel evaluatorsquality-diversity program synthesis Pareto archivefailure-driven test generation software repair held-out evaluationagent trajectory distillation workflow model routing cost qualityLLM evaluator calibration selection bias paired comparisonhuman-in-the-loop autonomous software change approval rollbackopen-ended self-improving coding agents archive empirical validationreward hacking specification gaming held-out tests AI systems05 / Primary sources
Research and standards behind the outline.
Primary papers and authoritative guidance, with the specific implementation lesson each contributes.
Evolutionary program search · 2025
Google DeepMind overview
AlphaEvolve
Pairs model-generated programs with automated evaluators and a program database that selects future prompt material. Strong evidence for the generate → verify → archive pattern when quality is objectively measurable.
Program search · Nature 2024
Read the paper
FunSearch
Combines a frozen language model, systematic evaluator, constrained program skeleton, parallel sampling, and a diverse island archive. It also states the key limitation: the method works best when evaluation is efficient and richly scored.
Self-modifying coding agents · 2026 revision
Read the paper
Darwin Gödel Machine
Empirically validates code changes, retains a growing archive of diverse agents, and explores multiple branches in parallel. The reported experiments explicitly used sandboxing and human oversight.
Metric-driven LM programs · ICLR 2024
Read the paper
DSPy
Represents language-model pipelines as parameterized computational graphs and compiles prompts or demonstrations against a declared metric. Useful precedent for separating workflow structure from optimizable parameters.
Software-agent benchmarks · ICLR 2024
Read the paper
SWE-bench
Frames repository changes as issue-resolution tasks evaluated in real execution environments. It demonstrates why repo-level loops need long-context inspection, multi-file edits, and executable tests rather than code-generation proxies.
Feedback memory · NeurIPS 2023
Read the paper
Reflexion
Uses feedback and linguistic reflection stored in episodic memory to improve later trials without changing model weights. This supports a trace-to-lesson loop, but does not replace external verification.
Evaluator reliability · ACL 2025
Read the paper
CalibraEval
Shows that model-based pairwise judges can change decisions with option order or identifiers. It motivates swapped-order checks, calibration, human audit samples, and uncertainty-aware promotion.
Risk management · NIST AI 600-1
Read the profile
Generative AI Profile
Organizes risk work around governance, content provenance, pre-deployment testing, and incident disclosure. It is the basis for keeping promotion, monitoring, and accountability outside the optimization loop.