07 · The runtime

The loop

One completion is a thought. An agent is a loop: think, act, observe, think again — until a stop condition fires.

In English

One model call is one thought. An agent is a loop: think, maybe use a tool, read what happened, think again — until it decides it is done, or you cut it off. Without a stop rule it will spend forever.

Try this: Click Step a few times. Notice the same model is called again with a longer transcript. Then hit the stop condition.
Keep this: An agent is a while-loop around a next-token engine.

Lab · the agent loop

Task: What is 15% of Tokyo’s population? Give a round number.

Think

Act

Observe

Stop

while

Step 1 / 7 · thought

I do not have a live census in weights I can trust. Search for Tokyo’s current population, then compute 15%.

thoughtI do not have a live census in weights I can trust. Search for Tokyo’s current population, then compute 15%.
toolweb_search
observationTokyo prefecture ~14.0 million (2025 estimate). Greater metro ~37 million. Sources disagree on “Tokyo”.
thoughtThe question is underspecified. Use the prefecture figure, flag the ambiguity, and compute 0.15 × 14,000,000.
toolcalculator
observation2100000
answerAbout 2.1 million people, using Tokyo prefecture (~14M). The greater metro is ~37M, which would be ~5.6M — ask if you meant that.

Pseudocode of every coding agent: while not stop: out = model(messages); if tool: messages += run(tool); else return out. The intelligence is in the model. The reliability is in the stop conditions, the tools, and what you append. That wrapper is the harness.

ReAct, “agent,” “assistant with tools” — same skeleton. Model, inspect the output, maybe run a tool, append, repeat. Stop on a final answer, a step limit, a spend limit, or a human.

The loop is where most production bugs live: infinite tool calls, lost observations, no stop condition, a transcript that never gets compacted. Getting the while-loop right is unglamorous and decisive.