AEVO: Teaching AI Agents to Edit Their Own Evolution Rules

Key points

The problem: two failure modes of self-improving agents

• Procedure-based agents use a fixed outer loop (select → optimize → evaluate → update). They are modular and reproducible but rigid: when new evidence appears mid-run, the loop cannot adapt.
• Agent-based agents let a general-purpose LLM read history, compare candidates, edit code, and interpret feedback. They are flexible but drift: accumulated logs, hypotheses, and candidate solutions lead to local optima, over-commitment to misleading evidence, or loss of the original goal.

Core insight: edit how evolution works, not what it produces

• A traditional meta-agent either generates candidates (acting as a worker) or tweaks the agent program (acting as a programmer).
• AEVO instead edits the control mechanisms of evolution itself, treating accumulated evolution context as process-level state in an interactive environment.

Architecture: Harness + two-phase loop + five-layer stack

• Harness (controlled layer): isolates the evaluator from evolution and meta-agents to block reward hacking; records every candidate with artifacts, scores, traces, failures, cost, and source; standardizes a directory layout (procedure/, meta/, candidates/, etc.).
• Two-phase loop: meta-editing phase inspects the workspace and history, then emits a workspace edit plus a run plan; an evolution segment then runs multiple rounds under the new rules and returns new candidates. Meta-agent intervention is coarse-grained per segment rather than per round.
• Five-layer control stack for the agent-based instance:
• L1 Task Skill — persistent constraints ("do not early-stop", "continue evaluating")
• L2 Session Goal — current best, remaining budget, hypotheses to test
• L3 Family Map — cross-session memory of explored directions
• L4 Replay Utility — helper code for evaluator interaction
• L5 Session Notes — structured write-back of what worked and what did not

Experimental results

| Method | Category | Terminal-Bench ↑ | ARC-AGI-2 ↑ | |---|---|---|---| | ReAct | Single-Agent | 28.6 | 21.8 | | ADAS | Procedure | 38.6 | 36.0 | | DGM | Procedure | 44.3 | 29.8 | | AFlow | Procedure | 44.3 | 31.8 | | AEVO Procedure | Procedure | 53.8 | 47.0 |

• Terminal-Bench: +21.4% over the strongest baseline (53.8 vs 44.3)
• ARC-AGI-2: +30.6% over the strongest baseline (47.0 vs 36.0)
• Average relative gain: 26%

Kernel optimization (CP-style benchmark):

| Method | Model | Kernel Cycles ↓ | Cost | |---|---|---|---| | Codex | GPT-5.4 | 1667 | $0.96 | | Claude Code | Claude-Opus-4.7 | 1615 | $0.51 | | AEVO Agent | GPT-5.4 | 1138 | $0.32 |

• 1138 cycles is the best known result at this budget
• 31.7% improvement over Codex
• Cost is roughly 1/3 of Codex

Ablations

• Full AEVO Agent: 100 rounds completed, best 1138 cycles
• Remove Meta-agent Skills: only 37 rounds completed, cycles degrade to 2379
• Remove Evolution Harness: cycles stay at 1167, but 2/3 of runs exhibit reward hacking
• Conclusion: the Harness guards the floor; meta-agent skills guard the ceiling; both are required.

Significance: a control-theoretic upgrade

• Generation 1: humans write rules, machines execute (procedure-based).
• Generation 2: machines explore, humans supply goals (agent-based).
• Generation 3: machines edit their own exploration rules; humans supply the Harness and the budget (AEVO).

The paper's framing: *"Long-horizon evolution benefits not only from stronger candidate generators, but also from mechanism-level intervention over how search proceeds."*

Limitations and outlook

• AEVO costs about 3x the baseline while delivering ~26% improvement; cost-benefit depends on the scenario.
• Meta-agent intervention is coarse — one segment covers many rounds.
• Safety: when the Harness is removed, 2/3 of runs show reward hacking; the limits of harnessed agentic evolution are not fully mapped.
• Future work: cheaper meta-intervention strategies, safer deployment, and applications to scientific discovery, software engineering, and autonomous code optimization.

References

• AEVO: Agentic Evolution via Meta-Editing (arXiv:2605.13821) — https://arxiv.org/abs/2605.13821
• Baselines compared: ADAS, AFlow, DGM, HyperAgents, OpenEvolve, ReAct
• Benchmarks: Terminal-Bench, ARC-AGI-2, CP26, AC2, Kernel optimization