How agents work

Jump to GitHub

saved run agent.py

No live model. Each step is saved data.

1Callmessages

2Responsetext or tool_call

3Executerun the tool

4Observeappend the result

A Messages

0 tokens

B Your program

C Program state

Architecture One model call

model Model returns the next message

returns

your program Your program sends the request

append

pane A Messages the list

run

pane C World tools and files

outside the list memory.db survives a new run

another loop Child agent own messages, own loop

call again with the updated list stop. no second call

Yellow is the current step.

The model chooses.
Your program acts.

An agent is a loop: call the model, run any tool call, append the result, call again. The model never sees files, networks, or databases. It sees the observation your program writes back into messages.

while True:

1 Call the model
Pass messages, and the tool schemas, into the API.
completions.create(messages, tools)
2 Read the response
You get text or a tool call. The model does not run the tool.
content or tool_calls
3 Execute
Your code runs the function. That is the only side effect.
functions[name](**args)
4 Observe
Append the tool result to messages. The next call can read it. Then call the model again, or stop if there was no tool call.
messages.append(tool_result)

Append the result. Call again. Stop when the response is text.

The model

Reads messages. Returns the next message. It does not open files, run commands, or write to a database.

Your program

Executes tools. Appends each result. That append is the observation. Then it calls the model again.

What you watched

One model call One API request. One message back. No loop.
One tool call A tool call is data. Your code runs the function.
The loop Call, run the tool, append the result, call again until the response is text.
Working memory The next call only sees the current messages.
Saved memory A file or database holds a fact after the message list is gone.
Multi-agent A second agent is a tool. The parent list only gets the returned string.
You choose You pick the next message. The program still runs tools and appends results.