DarwinXEvolving Agent Harnesses Through Natural Selection

Abstract

Method

• Selection over a population, not one lineage

  A run maintains an archive shaped as a tree, where each node is a harness snapshot with its edit delta, per-task scores, trial evidence, and distilled lessons. Nothing is discarded: a variant that loses overall may still hold the one edit that, combined with another branch's, unlocks a task neither solves alone. Parallel branches target different capability clusters, so the archive grows specialists with different solved-task signatures.

• A preserve-and-extend contract

  A child is admitted only when the measured evidence shows it solved something new without breaking what it already solved, which bounds what a win is allowed to cost elsewhere. When variants solve complementary tasks, their additive edits are merged, and the merged child is kept only if it covers the union of its parents' wins. Fitness is the benchmark's own verifier at avg@k — no gold solutions and no hand-picked winners.

• One edit interface for three signals

  Failure-derived diagnosis, teacher-derived demonstrations, and self-derived rollout contrast are all converted into harness edits rather than model-weight updates. The base model never changes, so every gain reported here is a statement about the harness alone.

Results at a glance

The four benchmarks progressively separate the evolution signal from the test: in-domain (Terminal-Bench 2.1), held-out task split (TerminalWorld), synthetic → real intent shift (WebArena-Infinity), and zero-shot cross-benchmark transfer (SWE-bench Verified).

Benchmark detail

WebArena-Infinity — synthetic → real intent shift

Evolution sees synthetic intents only; evaluation is the official 1,260-task real suite across 10 applications. We report audit-clean pass@1 throughout: trajectories flagged invalid by the validity audit count as failures even when the verifier passes them.

WebArena-Infinity audit-clean pass@1 by application. Click any column header to sort. External baselines are as-reported without re-auditing, which makes this comparison conservative for DarwinX. BU = Browser Use.

Application	Kimi	Qwen	Gemini+BU	GPT-5.5+BU	Monet (base)	Monet (DarwinX)	Δ
Elation clinical records	50.0	54.2	81.7	92.5	95.8	96.7	+0.9
Elation prescriptions	23.3	41.7	80.8	90.8	20.0	95.0	+75.0
GitLab plan and track	39.3	37.1	63.6	77.9	63.6	97.9	+34.3
Gmail	70.0	56.7	75.0	85.0	25.0	98.3	+73.3
Gmail accounts and contacts	40.0	33.3	61.7	87.5	21.7	91.7	+70.0
Handshake career exploration	50.0	50.5	50.5	83.5	36.5	84.0	+47.5
Linear account settings	54.2	65.8	73.3	81.7	43.3	94.2	+50.9
PayPal wallet	70.7	71.4	88.6	90.0	49.3	95.7	+46.4
Superhuman general	15.0	25.8	50.0	80.8	31.7	87.5	+55.8
Xero invoicing	52.5	55.8	80.8	93.3	39.2	96.7	+57.5
Overall	43.3	48.3	69.3	86.1	43.5	93.0	+49.5

93.0% audit-clean beats the strongest same-model baseline (GPT-5.5 + Browser Use, 86.1%) by 6.9 points and the top public agent (Gemini 3 Flash + Browser Use, 69.3%) by 23.7 points. Given differing action spaces across public agents, we treat the same-model gap as the primary comparison. In counts, the base solves 548/1,260 and the evolved harness 1,171/1,260.

Evolution as optimization

Screening score of every evaluated variant in the WebArena-Infinity run, in archive order. Hover a point for its score; the line is the best-so-far envelope.

Show every evaluated variant best so far

Anti-cheating audit

Validity audit on the 1,260 real tasks. Trajectories that reach the goal through the evaluation plane, privileged knowledge, or exploits are counted as failures.

Measure	Monet (base)	Monet (DarwinX)	Δ
Audit-clean pass@1	43.5	93.0	+49.5 pp
Confirmed invalid	23.5%	1.4%	−22.1 pp
Human review	5.1%	0.1%	−5.0 pp
Blocked attempt	14.3%	0.1%	−14.2 pp

Evolution does not buy the gain with shortcuts: it cuts invalid trajectories from 293 to 17, the evaluation-plane, privileged-knowledge, and exploit mechanisms disappear entirely, and the residual 17 are all raw-state mutations concentrated in a single application. Dropping unaudited successes too still leaves 92.9%.

Break down by

scale each row to its own total

Invalid trajectories before vs. after evolution. On a shared scale the second row is almost invisible, which is the result: 293 invalid trajectories become 17. Scale each row to its own total to inspect what the remainder is made of.

What the audit costs each harness

Each line runs from raw pass@1 to pass@1 after the validity audit. A long line means the harness was leaning on trajectories the audit rejects.

Show Monet (base) Monet (DarwinX) sort by audit loss

raw pass@1 base, audited DarwinX, audited

Ablation: what evolution changes

Anchored on Terminal-Bench 2.1, we compare base Monet (75.5%) with the evolved best (83.2%) along two axes: the skill-bundle diff between the two harnesses, and where on the benchmark the gains land. This is an exploratory attribution, not a per-skill causal ablation — the skills were co-selected, not independently randomized.

Limitations

• Transfer is measured in one direction, and is much smaller. A TB2.1-evolved harness reaches 84.2% on SWE-bench Verified, ahead of the fix-skill reference but inside a narrow band around it (official scores across the harnesses we compare span just 80.8–84.2%). The transferred gain is far smaller than the in-domain ones, and we make no in-domain SWE-V claim.
• Recombination's own contribution is not isolated. The archive and merge machinery can preserve and combine variants, but population search needs diverse wins before inheritance is useful, and the contribution of recombination relative to single-lineage mutation still requires controlled ablation.
• The claim is system-level, not per-operator. The experiments evaluate the complete system: archive, parent selector, recombination operator, and inference effort are not independently randomized, and public leaderboard rows use different models and effort settings. Matched-model deltas support a harness claim; any individual operator's contribution, and the attribution to a verification/contract mechanism, remain plausible rather than causal.
• TerminalWorld is small. With only 41 held-out tasks, one solve moves pass@1 by 2.4 points, and the matched Opus comparison (25/41 vs. 28/41, McNemar p=0.45) is suggestive rather than decisive.
• The validity audit is not a formal sandbox. Our WAI policy permits client-visible observation and semantic application operations but rejects privileged knowledge, evaluation-plane access, raw-state fabrication, database manipulation, and exploits. The static-plus-LLM audit is far stronger than a keyword heuristic, yet deeply dynamic construction can require human review and a small number of trajectories were unavailable, so we report both raw and conservative audit-clean scores.
• Freezing the base model is a choice, not a ceiling. It is what makes these deltas attributable to the harness. Coupling harness selection with weight updates, and treating the harness as an asset that outlives a base-model generation, are left to future work.

@article{zhang2026darwinx,
  title   = {DarwinX: Evolving Agent Harnesses Through Natural Selection},
  author  = {Zhang, Yifan and Dai, Yutong and Tan, Juntao and Yang, Luyu and
             Mullur, Rishi and Hoang, Thai and Hu, Zhiyuan and Zhu, James and
             Mui, Phil and Savarese, Silvio and Xu, Ran and Chen, Zeyuan},
  journal = {arXiv preprint arXiv:2608.07545},
  year    = {2026}
}