AITOPAI热报
PapersSelected

Google Paper Introduces Harness^2 for Self-Improving Agent Harnesses

elvis@omarsar0
Post
View the post
Why we picked it

The non-obvious finding is that recursive self-improvement can work without a verifier by letting candidate harness edits compete against each other, which sidesteps a common bottleneck in agent reliability. Builders of agent pipelines can test whether carrying evolved harnesses across tasks beats memory-based baselines like ReasoningBank in their own stacks. Readers can follow the linked arXiv paper and DAIR.AI's chat-with-paper tool to dig into the setup.

A Google research paper (arXiv 2610.14224) proposes Harness^2, a method for improving agent harnesses at test time without a verifier. Harness^2 generates several edits to the harness for the current task, tests them against each other, and keeps the best-performing combination, while lessons from those edits update a second harness carried to future tasks. It improved GLM-5 and Gemini 3.5 Flash in every tested configuration across WorkBench benchmarks WorkBuddy-Bench, JobBench and LAB. Running five candidate edits in parallel added another 2.9 to 3.9 points on WorkBuddy-Bench, while the ReasoningBank memory baseline fell up to 6.6 points below the base harness on coding tasks.

Original · elvis

Recommended, must-read paper on building recursive, self-improving agent harnesses for open-world tasks. DAIR.AI @dair_ai Learn to build custom harnesses, folks! Here is a highly recommended paper from Google on recursive agent harnessing. Great insights if you work on evolving your harnesses. It also offers an important discussion of recursive self-improvement and evolving harnesses for open-world tasks. They focus on improving your agent harness at test time without a verifier. This allows interesting scaling properties. Harness^2 proposes several edits to the harness for the current task, tests them against each other, and keeps the combination that works best. Lessons from those edits then update a second harness that the agent carries to future tasks. Results: It improved GLM-5 and Gemini 3.5 Flash in every configuration tested across WorkBuddy-Bench, JobBench, and LAB. Running five candidate edits in parallel adds another 2.9 to 3.9 points on WorkBuddy-Bench. ReasoningBank, a memory baseline, fell up to 6.6 points below the base harness on coding tasks in the same comparison. arxiv.org/abs/2610.14224 Chat with Paper: academy.dair.ai/papers/harness… 🔗 View Quoted Tweet 💬 8 🔄 3 ❤️ 16 👀 2026 📊 9 ⚡

#Harness^2#Google#agents#reasoning models#GLM-5

Keep reading

Recent stories with the same tag or section
PapersBuildersSelected17:16
Tencent Hunyuan releases WorkForge for training work agents

The dual programmatic-plus-semantic verifier design addresses the hardest part of agent training: rewarding long-horizon tasks without human grading. Teams building enterprise work agents can reuse the recipe to generate their own verifiable task suites from internal workflows rather than hand-crafted benchmarks. It sits alongside agent-environment efforts like task synthesis pipelines from other labs but anchors everything in expert-derived, evidence-backed tasks.

#Tencent Hunyuan#WorkForge#agents#open-source models
官方账号腾讯混元@TencentHunyuanSource
PapersBuildersSelected13:32
Meta and Michigan State paper finds LLM-built reasoning graphs unreliable

The non-obvious point is that most reasoning-graph pipelines assume the LLM's decomposition is correct without ever verifying it, so confidence scores inherit hidden graph errors. Builders of agent workflows and RAG-style decomposition systems should add graph-level validation before trusting computed confidence. The Qwen3-32B versus Qwen3-8B result is a concrete warning that scaling a model does not guarantee better sub-question answering.

#Qwen3#Meta#reasoning models#AI safety
rohanpaul_ai@rohanpaul_aiSource