OpenLoopEvolve: the Loop Policy is a versioned asset—Champion–Challenger before activation

Siqi Wang, Xinlin Li, Zhenglin Li, and Li Li (Tsinghua) treat the agent’s full control loop—not a prompt and not a skill library—as the object that must accumulate. OpenLoopEvolve (OLE) externalizes observation, planning, memory, action, verification, recovery, stopping, and budget control as a portable Loop Policy, packaged with applicability conditions and lineage into a versioned Bundle. Traces become attributable evolution evidence; an LLM proposes Challengers; Champion–Challenger paired evaluation under shared conditions feeds a robust release gate (benefit, evidence quality, tail risk, cost). Only gate-passing policies are released, and released policies activate only at a task boundary—never mid-task—so outcomes stay attributable; canary monitoring can roll back to the parent. Online mode triggers candidates from recent operational feedback; offline mode searches a candidate population from archived traces with multi-generation retention. On YC-Bench (year-long simulated business), with fixed DeepSeek-V4-Flash and official seeds, OLE-online and OLE-offline beat a fixed initial policy by +140% / +166% mean final funds, raise task success (~74%→88%/92%), improve annual survival (1/3→2/3 and 3/3), and cut mean max drawdown—while lowering average tokens per call even as runs survive longer. Code: yoyoshikc/OpenLoopEvolve.

Why it matters for code review

Self-improving overnight factories that rewrite the loop inside the same session are compounding unverifiable drift. Review harness-evolution PRs the way OLE gates releases: demand a named Loop Policy version, paired Champion–Challenger receipts on fixed conditions, task-boundary activation, and an explicit rollback path—reject “the agent updated its own CLAUDE.md mid-run.” Steady harnessed throughput compounds when control experience becomes a governable asset; FOMO for continuous self-edit without a promotion gate is the unconditional-accept drift Wei et al. already warned about. Pair with Wei’s Scale & Selection Champion–Challenger, Self-Harness’s held-out regression gate, and Macedo’s loop specifications when promoting overnight control-plane diffs.