Evidence-Governed Adoption of Self-Generated Updates in Autonomous AI Agents
Organizations: Independent Researcher
Abstract
Autonomous agents modify prompts, memories, tools, workflows, and policies using interaction histories. Yet a plausible modification is not necessarily suitable for persistent adoption: candidate selection is adaptive, evaluators may be imperfect, and local gains can hide degradation against an earlier baseline. We introduce Evidence-Governed Evolution (EGE), a framework for governing adoption of self-generated agent modifications. EGE separates candidate generation from confirmatory adoption, evaluates frozen candidates on fresh paired audits, requires meaningful improvement on a prespecified primary objective, protects declared capabilities through reference-relative non-inferiority tests, and applies governance constraints before bounded canary deployment. We establish an evaluator-calibrated simultaneous lower-bound construction for a countable sequence of adaptively generated candidates. It combines candidate-wise Hoeffding bounds with independent calibration samples comparing the operational evaluator with a prespecified reference procedure. We also formalise reference replacement through bridge tests, derive a cumulative degradation bound along a versioned reference lineage, and provide an audit-planning condition relating effect-size margin, sample sizes, and acceptance probability. A Monte Carlo stress test illustrates calibration under candidate-dependent evaluator optimism; it is a stylised statistical experiment, not an evaluation of a deployed agent. The guarantees are conditional on bounded paired scores, fresh sampling, a valid reference-evaluation construct, and declared evaluation distributions; they do not establish universal agent safety or robustness to arbitrary distribution shift.
Figures & tables
| Research stream | Primary question | Relationship to EGE |
|---|---|---|
| Reflection and iterative refinement | How can an agent improve a response or future attempt? | Can supply candidate modifications; does not by itself define persistent-adoption evidence. |
| Prompt and evolutionary optimisation | How can candidate prompts or configurations be searched? | Supplies candidate generation; development scores must not be reused as untouched confirmation evidence. |
| Safe policy improvement | When is a policy change supported relative to a baseline? | Provides related baseline-constrained reasoning; EGE targets heterogeneous agent-configuration updates and multiple protected capabilities. |
| Adaptive data analysis and confidence sequences | How can inference remain valid under repeated or adaptive analysis? | Supplies statistical tools; EGE instantiates a candidate-wise fresh-audit design and explicit calibration allocation. |
| Agent self-evolution benchmarks | Do self-generated experiences or harness updates improve agent capability? | Supplies evaluation settings; EGE governs whether a particular update is persistently adopted. |
| Canary deployment and release governance | How can a change be exposed gradually and monitored? | Supplies operational containment; EGE separates this from offline statistical evidence. |
| Bias | EGE harmful campaigns | Naive harmful campaigns | EGE power | Naive power | |
|---|---|---|---|---|---|
| 5,000 | 0.00 | 0/3,000 | 0/3,000 | 52.6% | 100.0% |
| 5,000 | 0.10 | 0/3,000 | 0/3,000 | 76.2% | 100.0% |
| 5,000 | 0.25 | 0/3,000 | 2,997/3,000 | 65.7% | 100.0% |
| 20,000 | 0.00 | 0/3,000 | 0/3,000 | 100.0% | 100.0% |
| 20,000 | 0.10 | 0/3,000 | 0/3,000 | 100.0% | 100.0% |
| 20,000 | 0.25 | 0/3,000 | 2,997/3,000 | 100.0% | 100.0% |
| Actual audit bias | Calibration bias | Harmful-acceptance campaigns | Observed rate |
|---|---|---|---|
| 3,000/3,000 | 100.0% |
| Global event index | Per-test | minimum | minimum | Total |
|---|---|---|---|---|
| 1 | 16,913 | 67,651 | 84,564 | |
| 10 | 31,650 | 126,597 | 158,247 | |
| 100 | 46,386 | 185,543 | 231,929 |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Meaning |
|---|---|
| Incumbent configuration before candidate | |
| Frozen candidate configuration | |
| Versioned stable reference configuration | |
| Experience available to candidate generation | |
| Prespecified primary-objective test set | |
| Prespecified protected-capability test set |
AI Agent Governance
AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.
See all 472 papers