AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Evolving the Agent’s Toolbox, Not Its Brain: DarwinX Improves LLM Agents by Selecting Better Harnesses

A large language model’s behavior is shaped by more than its weights. The surrounding harness—prompts, tools, memory, and control flow—can matter just as much, but letting agents rewrite that harness has proven fragile: an edit that fixes one task often silently breaks another, and early choices can lock in a mediocre path. Researchers at Salesforce AI Research say their new system, DarwinX, addresses this by treating self-improvement as natural selection over a population of harnesses while the model itself stays frozen.

DarwinX proposes small additive edits to the harness and admits a candidate only under a “preserve-and-extend contract”: the new version must improve on at least one task without regressing more than a small tolerance on what its parent already solved. Fitness comes from each benchmark’s own verifier, using average solve rate over k attempts (avg@k) rather than gold solutions. Failed variants are not thrown away; they stay in an archive so lineages that solved complementary tasks can later be merged. Proposed edits are driven by three evidence sources—failed rollouts, a stronger solver’s demonstrations, and contrasts between the agent’s own successful and failed attempts—all converted into harness changes.

The archive’s value is visible on TerminalWorld, a terminal-task benchmark. Four evolved specialists solved different subsets of 41 held-out tasks—24, 25, 26, and 27 tasks, respectively—while the merged harness solved 28, outscoring every specialist and every off-the-shelf agent tested. The authors say this shows why single-lineage hill-climbing is fragile: a globally weaker variant can hold an edit that another branch needs.

Across four benchmarks, the authors report an average gain of about 17 points. On Terminal-Bench 2.1, with GPT-5.5 frozen, DarwinX raised base Monet from 75.5% to 83.2% avg@5, a gain of 7.7 points, and reached 84.7% on the stronger GPT-5.6 Sol base. On TerminalWorld’s held-out split, it improved from 61.0% to 68.3% pass@1 (one attempt per task). On WebArena-Infinity, it evolved only on 300 synthetic intents, then scored 93.0% audit-clean pass@1—counting invalid trajectories as failures—on 1,260 real browser tasks, up from 43.5%, a 49.5-percentage-point jump, while invalid trajectories fell from 293 to 17. A Terminal-Bench 2.1 harness also transferred unchanged to SWE-bench Verified, resolving 84.2% of 500 issues, 3.4 points above a fix-skill reference.

The authors also list important caveats. Evolution is expensive: candidates must be re-run multiple times, so it is an offline job rather than a per-request update. The experiments evaluate the whole system, leaving the archive, selection rule, and recombination operator not independently isolated. They call the link between the gains and evolved verification/contract skills plausible rather than causal. TerminalWorld has only 41 held-out tasks, so one solve shifts the score by 2.4 points, and the one-task margin over the strongest off-the-shelf agent is not statistically decisive.

Still, the broader point stands: a frozen model need not be a fixed agent. Because harness edits are readable text and code, this kind of evolution may be easier to audit than weight updates. DarwinX does not prove that evaluation can replace training, but it suggests that selection over harnesses can turn ordinary evaluation compute into durable capability.