Exo paper: agents that rewrite their own harness at runtime, cutting model cost up to 82%
sarahcat21 · x · 2026-11-06
Researchers including Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica, Martin Casado and Ankur Goyal published "Exo: An Architecture for Self-Improving Systems at Runtime."
- Core idea: agent harnesses are usually hand-tuned or optimized offline; Exo collapses the optimization loop into task execution, letting agents revise their harness and strategies while working.
- Safety: a minimal protected mechanism layer preserves execution history and governs policy activation as the harness evolves through inspectable code changes.
- Results: on long-horizon objectives in an open-world game, Exo cuts model cost 12-53% vs Codex and Pi standalone and 45-82% on transfer runs at matched completion rates; on AppWorld/SREGym it cuts held-out model cost 35-61% with comparable correctness, trimming AppWorld mean task time 13% and SREGym median diagnosis time 35%; Continual Learning Bench coding cost falls 14% including learning.
Related event: Berkeley's Exo Lets Agents Rewrite Their Harness, Cutting Costs(2 posts)→
More from coding & agent
- Paper proposes Multi-developer Agentic Systems (MDAs) as the next scaling axis — karthik_r_n · 2026-11-06
- Codex agent runs 33 hours changing 115 files, fake 'GPT-6 Sol Ultra' meme — dscape · 2026-11-06
- Stagehand 4.4 Ships Crash-Safe Browser Sessions, Lands in LangChain Deep Agents — LangChain · 2026-11-06
- REINFORCE Zero: TII plus minimal IS ratio keeps async LLM RL stable for 20,000 updates, open-sourced — kastnerkyle · 2026-11-06
- Asking Codex to clean up scheduled tasks goes hilariously wrong — dscape · 2026-11-06
- Agent evals roadmap talk from AI Engineer NYC: video, slides, full guide — cwolferesearch · 2026-11-06