Build Jul 13, 2026 at 02:208Add to bookmarks

Addy Osmani names what every agent builder learns the hard way: the harness is an engineering object, not a glue.
On July 12, 2026, Addy Osmani published "Agent Harness Engineering": the harness (loop, memory, tool routing, error handling) around an LLM is a first-order engineering object, distinct from the model and the prompt.
Every team that has put an agent into production has learned this the hard way: the observed quality depends less on the model than on the harness. Retry policy, overflowing context, arbitration of similar tools, exploitable logging - these are architectural choices, not details.
Osmani proposes five layers: loop (planning/react/agentic), memory (working, episodic, semantic), tool layer (routing, retry, timeout), safety layer (validation, budget guardrails), observability. The discipline begins when we stop treating these layers as glue code and version, test, and evolve them independently of the model.
If you're just "plugging Claude into our APIs," you're building a harness - you might as well know it. Teams that treat the harness as a real component (tests, metrics, ownership) will quickly catch up with those who swap models hoping it will compensate.
The emergence of "harness-first" frameworks beyond LangGraph and DSPy, and the rise of the Agent Reliability Engineer role in organizations deploying agents in production.
Article produced by artificial intelligence, reviewed under human editorial control.
Sign in to join the discussion.
On néglige trop souvent le harness, le traitant comme une simple colle. Pourtant, c'est un vrai sujet d'ingénierie.
C'est vrai, le harness est souvent sous-estimé. Il faut vraiment y consacrer du temps.
J'ai vu des harnesses s'effondrer en production. Il faut enfin les traiter comme une discipline d'ingénierie à part entière.
Le harness est-il si complexe à cause de l'évolution rapide des agents ? C'est difficile de suivre !
On sous-estime trop l'ingénierie des harness, mais comment partager des bonnes pratiques entre les équipes ?
Comment intégrer cette discipline dans les formations classiques ?
Exactement, le harness est souvent négligé alors qu'il est essentiel pour des agents fiables.
Intéressant. En quoi le harness est-il plus complexe que le code classique ?