The agent loop, written by hand

An agent is a program that calls a model in a loop. Each iteration assembles the context (perceive), calls the model (reason), executes any tool calls the model requested (act), and appends the results to the history (observe). The loop ends when the model says it is done, or when your code decides it must stop: too many steps, too many tokens, too much time, no progress, or a human interrupt. The model decides what to do next; your code decides whether it is allowed to keep going. Every framework and platform you will be handed is a variation on this loop, and most production incidents with agents trace back to a missing stop condition, unbounded context growth, or an error that was thrown instead of observed.

Key takeaways:

• The loop is ordinary code you own. Stop conditions, budgets and state are your responsibility, not the model's.
• Input tokens grow every step because the whole history is resent. Cumulative input cost grows roughly with the square of the step count.
• Tool failures should come back to the model as observations it can act on. Exceptions end the loop; observations let it recover.
• "The model said it was done" is only one of several stop reasons. Always record which one fired.
• Instrument every step: model call, tool call, tokens, latency, stop reason. Without the trace you cannot explain a failure to the customer.