Agent = Model + Harness: Google engineers show how failing tests rewrite the system prompt
Roger_M_Taylor · x · 2026-09-15
A long thread distilling agent harness engineering, built on the thesis Agent = Model + Harness.
- Google engineers let the AI rewrite an agent's system prompt against failing tests: a failed check becomes the next repair job, and a regression suite watches for whatever that repair breaks — even instructions become testable and improvable
- The compiled page maps the system into six parts: guides (rules, constraints, lessons), sensors (tests/linters/validators), loop (run-check-retry-escalate), memory (state, artifacts, decisions across runs), permissions (tool/write approvals), observability (what happened, what it cost, where it failed)
- Quoted article "From 296 Agents to One Graph": 296 agents mean 296 policies and failure modes; a graph + loop is the better structure
Key shift: agent engineering is about the harness, not the model.
More from coding & agent
- Models don't have agency, systems do: the one-token trick that seeds tool use — sethjuarez · 2026-09-16
- GPT-6 Astra demos: playable games, robot demos and 3D pipelines built in hours — socialwithaayan · 2026-09-16
- Proposal: hand predictable UI workflows to a local runner to cut computer-use agent costs — Iinganouo · 2026-09-16
- Tutorial: Tracing agent invocations and LLM calls with OpenTelemetry — dl_weekly · 2026-09-16
- LangChain makes every Managed Deep Agent an MCP server callable by coding assistants — LangChain · 2026-09-16
- LlamaIndex on Building SDKs with Stainless, Whose Team Joined Anthropic — llama_index · 2026-09-16