Early-Signal Detector

method · first seen 4 Oct 2026

agent harness evolution

Also: EMHO, llm-harness, agent harness, agentic harness, harness evolution, LLM agent harness, coding agent harness, harness optimization

Iteratively optimizing the surrounding scaffolding of an LLM agent using execution feedback.

Agent harness evolution treats the code around a fixed LLM (context construction, state, tool dispatch, verification, memory) as the object to optimize, using iterative proposal-and-evaluation loops driven by execution feedback. The sources show a cluster of methods papers: MESH-Harness recombines modular harness components with bandit-guided search, a causal improvement graph replaces proposer-centric meta-harness search, Self-Improvement Memory carries harness edits across benchmarks and models, and REINS applies a self-evolving harness to real-time trajectory planning. A Berkeley paper with Ion Stoica studies harness-side cost optimizations. GitHub repos from Didi, Reflection and individual developers build customizable coding harnesses. The term 'harness' itself is also used loosely for agent runtimes in general.

If model weights stay fixed while the scaffolding is searched automatically, much of agent performance and cost tuning could move from manual prompt and tool engineering to reproducible optimization loops. The Berkeley result that harness changes mainly affect cost, not accuracy, is a useful check on how much these methods can deliver.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

01419 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 11 on 16 Nov 2026 · 57 mentions in 28 days · 57 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration -1.07

Earliest sources

Latest sources

Mentioned together with

recursive self-improvementvision-language-action modelmulti-objective optimizationmulti-agent LLM systemcoding agentscausal improvement graphlinear upper confidence bound banditcoordinate ascentterminal agentlow-rank adaptation