The agent loop

The agent loop

The loop is the smallest useful mental model for an agent. A request enters the runtime with a goal and state. The model decides whether it can answer or needs another action. If it requests a tool, the application validates and executes it, then feeds the result back into the next step.

Stopping conditions

Every loop needs explicit stopping conditions: task completed, maximum steps reached, budget exceeded, policy denied, tool failure, or human escalation.

Why loops fail

Unbounded loops often come from ambiguous success criteria, tools returning low-quality results, or the model repeatedly choosing the same action. Add step budgets and detect repeated tool calls.

Example

Apply The agent loop to a realistic production scenario and trace the decision step by step.

Code

Add implementation notes or a runnable example for this concept.

Practice

Write down the inputs, expected output, constraints, and one failure case for The agent loop.

Challenge

Design a production-ready solution for The agent loop and explain one important trade-off.