AIAny
Icon for item

DarwinX: Evolving Agent Harnesses Through Natural Selection

Treats agent self-improvement as natural selection over a population of harnesses (prompts, tools, skills, control flow), evolving a frozen-model agent by selecting harness edits that extend capability without regressing others. Uses a preserve-and-extend contract, lineage archive, and verifier-driven fitness (no gold solutions) to recombine complementary edits and transfer across benchmarks.

Introduction

Most agent self-improvement loops edit prompts, tools, or control flow but do so along a single lineage that can plateau or silently regress other skills. DarwinX reframes harness design as an evolutionary selection problem: keep model weights frozen and search a population of harness variants, admitting children only when they add measurable wins without giving up existing ones and retaining alternative lineages for later recombination.

Key Findings
  • Empirical lifts across four benchmarks: an average gain of ~17 points from harness evolution alone; Terminal-Bench 2.1 improved to 83.2% (and to a verified frontier of 84.7% on a stronger base), TerminalWorld held-out hit 68.3%, WebArena-Infinity pass@1 rose from 43.5% to 93.0% audit-clean. These show gains transfer across tasks and verifiers.
  • Preserve-and-extend contract quantifies net gain vs. bounded regression so each accepted edit must extend coverage without large regressions, preventing local-win regressions common in single-lineage edits.
  • Archive-shaped population lets complementary specialists from different lineages be recombined; nothing is discarded so rare but useful edits survive for later inheritance.
  • Fitness is benchmark-driven via each benchmark’s verifier (no gold solutions, no hand-picked winners), so improvements reflect measured task competence and can generalize across verifiers and base models.
Who It's For and Trade-offs

Great fit if you need durable agent capability improvements without retraining model weights, especially for LLM-driven agents where prompts, tools, memory, and control flow dominate performance. It suits research teams or engineering groups that can run repeated verification trials and want transfer across benchmarks. Look elsewhere if you lack compute budget for repeated evaluation loops, if your tasks lack reliable verifiers, or if you require model-weight upgrades (DarwinX never moves weights and therefore cannot change raw model capacity).

Where It Fits

DarwinX sits between ad-hoc prompt/tool tuning and full model fine-tuning: it treats harness search as selection rather than gradient-based training, addressing path dependence and cross-task interference that hamper single-lineage self-editors.

How It Works (brief)

The method proposes harness edits from failure-, teacher-, or self-derived signals, scores candidates by per-task solve rates (avg@k) under each benchmark’s verifier, and promotes only those children that improve some tasks while keeping regressions within a bounded threshold. An archive records edit deltas and evidence so lineages accumulate capability and can be recombined later.

Information

  • Websitearxiv.org
  • OrganizationsAffiliation: Salesforce AI Research, Affiliation: Salesforce Agentforce
  • AuthorsYifan Zhang, Yutong Dai, Juntao Tan, Luyu Yang, Rishi Mullur, Thai Hoang, Zhiyuan Hu, James Zhu, Phil Mui, Silvio Savarese
  • Published date2026/07/31

Categories

More Items

Supports multimodal scientific understanding, long-horizon agentic workflows and scientific tool interaction using a unified pipeline of multimodal pretraining, supervised fine-tuning and scalable multi-task reinforcement learning. Distinctive features include time-series modules for signal forecasting and a separate Memory Decoder that enables rapid domain specialization without changing the frozen 397B backbone.

Evaluates whether LLM-driven storyteller agents preserve long-horizon logical consistency under adversarial player interventions. Introduces NCP-Bench (100 movie-derived narrative environments) with structured trajectory/commitments and automatic violation checks; finds strong LLMs often contradict themselves across multi-turn interactions.

Uses a stronger 'builder' model at inference time to construct executable harnesses that boost weaker target models without parameter updates, mainly by turning unstable reasoning into deterministic code, routing, and strict answer-format enforcement.