Lesson 16 — Observations back into the conversation

The model only knows what you tell it

An observation is the result of a tool call, written back into the message list. The model has no other way to learn it: if you run the tool but forget to add the result, the next call looks exactly like the last one — and the model asks for the same tool again.

So each step adds two messages:

1. the model's tool call, as an "assistant" message;
2. the tool's result, as a "tool" message (with the tool's name).

Real APIs do the same with small differences: OpenAI uses "role": "tool"; Anthropic puts a tool_result block inside a "user" message.

Why it matters: "the agent keeps repeating itself" is one of the most common agent bugs, and a missing observation is usually the cause.