Google open-sources Harness², a 24.6k-star framework for recursively optimizing agent harnesses around frozen models

adnan_hashmi · x · 2026-11-20

Google Research released Harness² (google-research/harness2), a plug-and-play framework for recursive agent harnessing in open-world settings, already at 24.6k stars and 2.5k forks under Apache-2.0. Core idea: a harness is everything around a model that makes it an agent (prompts, tools, memory, skills, orchestration), and harnessing is optimizing that system while keeping model weights frozen. It links two levels of recursion: task-level recursion refines the execution harness for the current task via a contrastive Propose–Probe–Compose step—propose harness edits, probe their effects on the same task, compose a refined harness. The accompanying diagram shows the reference architecture: context construction (system prompt, guardrails, memory, skills), middleware (hooks on every tool call with plugins, tool guidance, scripts), and orchestration (subagent delegation), decomposed into components C1–C8.

Original post →

More from coding & agent

coding & agent channel →