Cohere Labs talk: Harness² recursively improves agent harnesses at test time
Cohere · youtube · 2026-11-14
Cohere Labs Open Science Community hosts Ruiyao Xu (PhD candidate, Northwestern) presenting "Harness²: Recursive Agent Harnessing for an Open World" — improving an agent's harness via recursion while keeping model weights fixed.
Two connected levels:
- Task-level recursion improves the execution harness (prompts, tools, memory, skills) via contrastive Propose–Probe–Compose steps: propose edits, probe their effects on the current task, and compose a refined harness from behavioral evidence, exploring candidates in parallel or sequentially.
- Harnessing-level recursion improves the improvement process: after each task, adopted harnesses and evidence-backed lessons are retained in an evolving library guiding future refinement.
The talk covers how the levels connect within-task improvement with cross-task learning, plus experiments showing open-world agents adapting at test time without ground-truth verifiers.
More from coding & agent
- Qwen 3.5 4B scores 10/10 running a coding agent fully offline on a 16GB MacBook Air — Ok-Brush8377 · 2026-11-14
- Ditching $500/Month SaaS Stack: One MCP Lets Claude Code Run Outbound for 91 Cents — aryanXmahajan · 2026-11-14
- FastMCP 4.2 Ships Skills Over MCP, Pairing Tools With Usage Guidance — jlowin123 · 2026-11-14
- Free guide: build a one-person agency powered by a team of AI agents — EXM7777 · 2026-11-14
- Dev reverse-engineers Waymo's private API into a Cloudflare Worker MCP for Claude — MicahBerkley · 2026-11-14
- Dev's fix for chaotic agent work: a feed-shaped todo list with action buttons — _rchaves_ · 2026-11-14