내일 정독할게요
Google's team dropped a 9-page PDF on Harness Engineering. prompt engineering got replaced by one formula:
Agent = Model + Harness
the catch: same Claude Sonnet, same benchmark. only the harness changes
the playbook runs six steps:
step 1 -> add guides
AGENTS.md, rule files, constraint docs. every line in there is an old failure that can't happen twice
step 2 -> add sensors
linters, tests, validation scripts. the agent grades its own output before a human ever opens it
step 3 -> build the agentic loop
plan, execute, verify, fix. retries capped, budget capped, and a stuck agent hands the problem up
step 4 -> externalize memory
the model wakes up blank every session. the harness holds the state, the decisions and the artifacts
step 5 -> enforce permissions
which tools, how many writes, what needs a human. safety lives in the harness, not the prompt
step 6 -> wire observability
every tool call, every dollar, every retry gets logged. a trip wire fires the moment behavior drifts
the result: the agent stops being a demo and starts being infrastructure
each failure becomes a permanent fix, so the next run opens smarter than the last
that's the distance between an agent you show off and one a client pays you to leave running
one harness buys you a reliable loop. wire several together and you get graph engineering, where one agent's failure becomes another agent's sensor
the PDF ends at the harness. the graph is the layer after it
then read the article below ↓
1
100








