DarwinX: Evolving Agent Harnesses with Frozen Model, +17 Points on Average
daniel_mac8 · x · 2026-08-15
A new paper introduces DarwinX, which evolves agent harnesses (prompts, tools, skills, control flow) through natural selection while keeping model weights frozen. Across four benchmarks, it improves performance by 17 points on average: Terminal-Bench 2.1 rises to 83.2% (matched base) and 84.7% (stronger base); TerminalWorld held-out reaches 68.3%, ahead of off-the-shelf agents; WebArena-Infinity pass@1 jumps from 43.5% to 93.0%; and the harness transfers to SWE-bench Verified. The authors emphasize that general agent competence evolves, not benchmark-specific patches.
Related event: Salesforce Releases DarwinX to Evolve Agents via Natural Selection(3 posts)→
More from coding & agent
- NVIDIA NeMo Switchyard Enables Routing AI Agents Across Models — NVIDIAAI · 2026-08-15
- NVIDIA open-sources NeMo Switchyard for dynamic model routing in agent workflows — NVIDIAAI · 2026-08-15
- Sleek Agent Skills Enable AI to Design Mobile App Screens — tom_doerr · 2026-08-15
- Gemini 3.7 Flash leads Computer-Use Arena ahead of Claude Fable — cgarciae88 · 2026-08-15
- Every Engineer Shares 4-Layer Defense for AI Employee Safety — every · 2026-08-15
- Agent Observations: Cracking CAPTCHAs, Guessing Passwords, and Inefficiency — DhruvBatra_ · 2026-08-15