this harness paper is f*cking gold
Anthropic engineers laid out what actually makes Claude Code work so well. spoiler: it's not the model.
the secret: the harness. that's the entire gap.
prompt engineering → context engineering → harness engineering.
the 6 layers of the harness:
layer 1 → tools
fewer tools, one job each.
every error tells the agent what to try next.
layer 2 → context
short claude-md → files on demand → compaction when it fills up.
layer 3 → permissions
reads + tests → auto.
deploys + deletes → ask a human.
everything → sandbox.
layer 4 → hooks
before a tool runs → block secrets and protected files.
after every edit → format, lint, typecheck. no exceptions.
layer 5 → verification
fresh reviewer reads the diff, not the summary → runs the tests → fail? fix → rerun.
layer 6 → evals
every harness change scored → success, tokens, turns → keep what wins.
the model is the engine. the harness is the whole car.
they even packed all 6 layers into one prompt. paste it into Claude Code and it engineers the harness for you.
6 layers. 1 prompt. 0 infinite loops.
then read the article below ↓
Esto es una locura. Google acaba de sacar un curso gratis de 2 horas sobre Harness engineering completo, del primer paso al ultimo:
0:29 - Harness engineering desde cero
11:08 - tu primer agente, bien construido
28:08 - Eval engineering con un agente “Chief of Staff”
1:00:13 - Loop engineering
1:23:19 - un sistema de grafos que se mejora solo y nunca se detiene
He visto cursos de 500$ que no tienen el valor que tiene este video
this harness paper is f*cking gold
Anthropic engineers revealed what actually makes Claude Code work so well. spoiler: it's not the model.
the secret: the harness. that's the whole gap.
prompt engineering → context engineering → harness engineering.
the 6 layers of the harness:
layer 1 → tools
fewer tools, one job each.
every error tells the agent what to try next.
layer 2 → context
short CLAUDE.md → files on demand → compaction when it fills up.
layer 3 → permissions
reads + tests → auto.
deploys + deletes → ask a human.
everything → sandbox.
layer 4 → hooks
before a tool runs → block secrets and protected files.
after every edit → format, lint, typecheck. no exceptions.
layer 5 → verification
fresh reviewer reads the diff, not the summary → runs the tests → fail? fix → rerun.
layer 6 → evals
every harness change scored → success, tokens, turns → keep what wins.
the model is the engine. the harness is the whole car.
they even packed all 6 layers into one prompt. paste it into Claude Code and it engineers the harness for you.
6 layers. 1 prompt. 0 infinite loops.
full breakdown in the Harness Engineering guide below ↓