[research] DarwinX: Evolving agent harnesses adds +17 pts without touching model weights

[github-actions[bot]]

🔬 The Finding

Salesforce AI Research released DarwinX, a framework that treats agent self-improvement as natural selection over a population of harnesses — the prompts, tools, skills, and control flow wrapped around a frozen LLM. Unlike single-lineage self-edit loops (which tend to regress on tasks they weren't optimized for), DarwinX uses a "preserve-and-extend" contract: a variant only survives if it extends coverage without regressing elsewhere. One evolution loop added ~17 points on average across four benchmarks, including WebArena-Infinity real-task pass@1 jumping from 43.5% → 93.0%.

⚙️ What It Means for Agentic Workflows

1. The harness is the lever, not the model. Teams building automated GitHub workflows can likely get outsized gains by systematically evolving their prompt scaffolding, tool selection logic, and control flow — no fine-tuning or new model required.
2. Multi-lineage archives beat greedy editing. If you're iterating on an agent harness, maintain alternative lineages rather than always taking the "best" single update — the archive enables recombination that escapes local optima.

🔗 Source

DarwinX: Evolving Agent Harnesses Through Natural Selection — 31 July 2026

Generated by Daily Agentic AI Research Digest · 96 AIC · ⌖ 11.2 AIC · ⊞ 24.2K · ◷

• expires on Aug 24, 2026, 9:49 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-08-24T09:49:17.976Z.

Closed by Workflow