EnvOpt Paper: Optimizing Agent Environments Cuts Cost 72.7% and Boosts Pass Rate 14.34%
dair_ai · x · 2026-11-06
Researchers from UC Berkeley, University of Chicago, and Bespoke Labs propose EnvOpt, automating a third axis of agent improvement beyond parameter tuning and harness optimization: the stationary environment agents operate in (e.g. a company Slack workspace, docs, tickets).
A coding agent runs the target agent on training tasks without graders or reference answers, iteratively edits how the environment is organized and presented, and stops based on a validation set.
Key results:
- On AutomationBench, cuts Claude Haiku 4.5's cost by 72.7% and raises pass rate by 14.34%
- The same optimized environment transfers to Opus 5 with 68.1% lower cost and a 1.34% gain
- Stacked on Meta-Harness's optimized harness, it lowers cost another 83.4% and adds 11.17% accuracy
The paper argues that how information is represented to an agent shapes both cost and success, so environments deserve systematic optimization, and documents which environment-editing strategies succeeded or were rejected.
More from coding & agent
- Harness adapters: per-agent-framework LoRAs cut tool calls, $5 for 1,350 eval sessions — LysandreJik · 2026-11-06
- Agents wrote a polished F1 champions site via multi-model rounds, cloning a Deimos-style build — evilrabbit_ · 2026-11-06
- Open-source local MCP reads your WhatsApp/Telegram and drafts replies in your tone — SuccessfulCress7441 · 2026-11-06
- 10 Open-Source Repos That Level Up Your Coding Agent, Top One at 313k Stars — victor_explore · 2026-11-06
- Team beats Mem0 on its own benchmark after 6 weeks, then hits a 7x-too-big submission bug — gauzzastrip · 2026-11-06
- Repurposing a Logi 9-key macro pad as physical YES/NO/CONTINUE buttons for coding agents — flngr · 2026-11-06