Observation
After an action, an agent needs a trustworthy observation of what actually happened before choosing the next step.
After an action, an agent needs a trustworthy observation of what actually happened before choosing the next step.
What the agent attempted.
What the environment returned.
How state should change.
Understand the working model
An observation should come from the tool or environment, not from the model imagining that an action succeeded. Store structured outcomes such as status, IDs and error codes when possible.
This prevents phantom progress. If an API call failed, the next reasoning step must see that failure instead of assuming the record was updated.
Practice
Apply the pattern above to one real task you already do. Keep the first version small enough that you can inspect every input and output.
Do not let the model write its own “success” observation without checking the real tool result.
Every action should produce a real observation that updates state before the agent reasons again.