Exo: Berkeley's agent runtime that edits its own harness, cutting model cost 35-61%
dair_ai · x · 2026-11-06
Researchers from UC Berkeley, Andreessen Horowitz, and Braintrust present Exo, an agent runtime architecture that moves harness optimization from an offline loop into normal task execution.
Key idea
- Traditional harness optimization needs representative datasets and many expensive runs; 75% of interviewed teams evaluate agents without benchmark sets.
- Exo lets the agent edit its own harness code while working. A minimal protected mechanism layer preserves execution history and controls which policy version is active, making every change an inspectable code diff.
Results
- On AppWorld and SREGym, Exo cuts held-out model cost by 35-61% vs. the same starting policy without self-editing, at comparable correctness; AppWorld mean task time drops 13%, SREGym median diagnosis time 35%.
- On standalone Codex and Pi runs, model cost falls 12-53% (45-82% in transfer) with matched task completion.
- On Continual Learning Bench coding tasks, total model cost including learning drops 14%, with both policies resolving every issue.
Related event: Berkeley's Exo lets agents rewrite their own harness at runtime(4 posts)→
More from coding & agent
- Click-to-summarize UI built on DocETL rewrite directives tames Anthropic's 421-page S-1 — sh_reya · 2026-11-07
- Amortized Reasoning Index: a retriever that self-improves during search, no labels needed — lateinteraction · 2026-11-07
- Linear token pricing breeds bloated agents — rationalized harnesses could run at ~300k tokens — JoshPurtell · 2026-11-07
- Linear token pricing kills incentives: optimized agents could run at 300k tokens — JoshPurtell · 2026-11-07
- Plug your phone into Codex: delete bloat apps, clear voicemail, tame old Kindles via ADB — angadsg · 2026-11-07
- 'Software factory' debate: marketing fluff or necessary infra for scaling coding agents? — vikvang1 · 2026-11-07