66-page survey maps 179 papers on agent harness self-evolution and eval rigor
udmrzn · x · 2026-10-24
A 66-page survey on Harness Self-Evolution (401 GitHub stars, bilingual) systematically maps self-evolving agent routes from Prompt, Memory, and Skill to Workflow and executable code, curating 179 papers as a literature map. Its key concern: agents change prompts, skills, and code all at once and scores rise — but which change actually worked? Most self-evolution experiments can't answer. The survey's eval discussion offers three checks: did the change take effect, is the gain attributable to it, and does it hold on other tasks. If any answer is no, don't trust the improvement.
More from coding & agent
- Polylane Launches for Cloudflare: Proactive Agents Watch Workers, Merge ~8 PRs a Day — dinasaur_404 · 2026-10-24
- Factory tested Reflection's Beam pre-launch via Together AI for day-0 support — togethercompute · 2026-10-24
- Kimi K3.1 vs GPT-6 Astra on the same agentic 3D task: full token and cost breakdown — GapNew4766 · 2026-10-24
- Benzi: compiler-backed coding agent hits 82.4% SWE-bench Verified for under $13 — DonkeyTheKing · 2026-10-24
- Bench shows GPT-5.6/6 Luna crush multi-agent coordination while DeepSeek V4 fails — JoshPurtell · 2026-10-24
- Payman AI launches MCP server for natural-language payments and payee management — modelcontextprotocol · 2026-10-24