Berkeley Paper: Harness Tweaks Cut Agent Costs Up to 2.6x Without Changing Models
dair_ai · x · 2026-11-11
A UC Berkeley paper, 'System Optimizations for Cost-Efficient Agent Harnesses,' delivers a counterintuitive finding: on the same model and task, harness-side changes like prompts and tools barely move accuracy but swing execution cost by up to 2.6x. To cut agent bills, change the harness, not the model.
- Where cost goes (from agent trace analysis): reading existing context is the biggest source — up to 73% of total cost even with caching; tool calls add up to 26x cost at the same token count; only 27% of tool outputs are actually used; many turns go to filesystem exploration.
- Agency platform: The team built a harness-interposition layer that sits between existing harnesses (Claude Code, Codex) and the model, so an optimization written once works across harnesses.
- Results: Five proof-of-concept techniques cut costs up to 35% with marginal accuracy loss. If your agent bill keeps growing, audit repeated context reads and extra tool calls before switching models.
Related event: Berkeley Paper: Agent Harness Choice Can Cut Costs 2.6x(2 posts)→
More from coding & agent
- A/B your agent harness: score accepted results per dollar on hidden cases to know if improvements earn their cost — krishnan · 2026-11-11
- Flex-RSI preprint breaks down self-improving agents: editable harness, frozen weights, one-shot hidden test — krishnan · 2026-11-11
- Merge claims 2.9x more Slack tools and full error recovery vs Slack's MCP server — shensi · 2026-11-11
- Google DeepMind ships Gemini 4 Carbon, which ports 90s game Theme Park to web in one shot — demishassabis · 2026-11-11
- The Doherty Threshold: IBM's 400ms UX latency rule modern OSes like Win11 still violate — lauriewired · 2026-11-11
- Claude Code silently uploads local project docs to online Artifact, user flags privacy risk — OwariDa · 2026-11-11