Google open-sources Harness², an agent that rewrites its own setup per task
Arindam_1729 · x · 2026-11-11
Google open-sourced Harness², an agent that rewrites its entire harness (system prompts, guardrails, memory, tool instructions, subagents, skills, plugins) for every new task — without touching model weights.
How it works:
- Propose: the agent suggests changes to its own harness
- Probe: it tests multiple versions against the same task
- Compose: it diffs those runs and combines what worked
Key twist: no answer key or external grader needed — it improves by comparing its own runs.
The author also cites DeepSeek's dsh-superagent prototype as evidence that the ceiling on multi-task coding agents is human attention, not the model: a router session triages tickets into isolated workers, and unrequested work starts in observe mode.
More from coding & agent
- Plurai launches free agent simulation testing inside Claude Code and Cursor — PrajwalTomar_ · 2026-11-11
- AI-built game nerfed 40% after Claude playtest bot beat it in 6 minutes — builderjaydub · 2026-11-11
- Anthropic launches Claude Tag for Slack, claims 65% of its product code is AI-written — nikola_mr64990 · 2026-11-11
- The key to expert-driven evals: maximize information gain per unit of expert attention — danshipper · 2026-11-11
- Aqua launches Notetaker: meeting notes piped to AI agents via MCP — Scobleizer · 2026-11-11
- KAIST's WikiTrace: evidence-linked memory fixes agent "memory-induced degradation" — dair_ai · 2026-11-11