Agentic Loop

The agentic loop is the control structure that turns a text-predicting model into something that gets work done: the model produces a turn; if that turn contains a tool call, the host executes it and appends the result to the conversation; the model runs again with the new information; repeat until it produces a final answer or a limit is reached.

Everything characteristic of agents follows from this loop. The model can investigate — read a file, run a search, look at the result, and choose the next step based on what it found. It can also fail in loop-specific ways: repeating an action whose effect it cannot observe, or terminating early because it believes work is done.

Practical design centres on a few pressures.

Observation must reflect reality. If a write goes to one place and the corresponding read comes from another, the model sees its own change fail to appear and does it again — producing duplicates. Read and write paths must agree; see Draft and Published.

Every iteration consumes the Context Window, so long loops need artifacts stored outside it and referenced by handle.

Limits are real. Step caps and request timeouts wrap the whole turn including tool execution, so a slow tool can abort a turn before its own honest report is ever delivered.