Smarter harness is as important as smarter model #ai #aiagents ##programming #claude #learntiktok

The smartest AI agents are not running smarter models. They are running with smarter harness. Take an example of claw deep research. Instead of giving one agent the full deep research task, it spawns sub agents. Each one gets a narrow slice. Research this angle, summarize this document, check this source, and runs in its own isolated context window. A supervisor agent collects their outputs, reconciles contradictions, and produces the final answer. The model never changes, the architecture does, the numbers are from Anthropics own benchmarks, a flat single agent setup processing the same task in one pass scores around 40%. The sub agent supervisor setup, same model, same task scores 90. That is a 90.2% relative quality gain. The cost is roughly 15 times the token count because every sub agent carries its own full context plus the coordination overhead. This is why every deep research feature you have used feels qualitatively different from a normal chat. It is not a smarter model. It is a fan out of sub agents doing parallel exploration. Each one staying focused inside a small context window, all of them feeding a coordinator that argues with itself before answering you. Three rules if you want to use this in production. One, do not spawn sub agents for tasks a single agent handles well. The 15 times cost is wasted on simple queries. Two, give each sub agent a single narrow objective. vague briefs collapse the quality gain. Three, design the supervisor's reconciliation step deliberately. The whole pattern fails if the coordinator cannot detect contradictions between sub agent outputs. The lesson is not that more agents is always better. It is that on hard tasks, parallel narrow attention beats a single wide one. Pay the tokens when the task earns it. Its complete breakdown is in my school community. Link is in bio.