Decoding Claude: E01 - The Loop
Imagine that you hire one person to do all the work at your company. They understand what you want to get done, plan tasks, execute them, decide which tools they want, use those tools with their limited knowledge and assess the output themselves. This person is simply doing too much.
An AI agent that gets a task but no framework for delegation faces the same dilemma. It lacks the power to give its full attention to each aspect of the job. When a task is too complex, it concentrates on understanding and delivery, and fares poorly at execution.
In The Office, Michael is the one who decides things. Dwight is the one who does things. Michael doesn't book his own travel, research his own talking points, or chase down his own paperwork. He tells Dwight what needs to happen and Dwight executes. Michael's job is judgment. Dwight's job is action. That's your loop.
The Core Loop
We code this separation of planning and execution into the core loop. The top level LLM can act as a Project Manager. It delegates tasks to experts in the form of sub-agents or specialized tools. The PM concentrates on understanding the task, breaks it into sub-tasks, hands it to experts, and assesses the output. Once the output fulfills the query from the user, it provides the answer.
Everything else in this series is just this loop, with more people in the room.
Next: Decoding Claude: E02 - More Desks, Same Boss.
New here? Start with Decoding Claude: How Agentic AI Actually Works — where the series began.
References
Learn Claude Code (Github) - The Agent Loop: Link
