Observation
Whatever the agent perceives from the world after it acts — the feedback it reasons over next.
Think of it like
Looking up after each chess move to see how your opponent responded before planning the next one.
Example
The agent clicks “submit,” and its observation is the confirmation page — or the red error banner telling it something went wrong.
How it actually works
Observation is the perception half of the loop: tool results, screenshots, API responses, error messages — the environment’s reply to the agent’s action. Grounding each step in fresh observations is what separates an agent from a model spinning stories in a vacuum. The catch is that observations can be noisy, huge, or adversarial, and the agent has to read them correctly to act well.
For product teams
Agents are only as smart as what they can actually see after each move.
For engineers
The environment’s response to an action, fed back into context as the input for the next reasoning step.
Related
- Tool Result — The tool version of an observation.
- Action Space — The set of moves the agent chooses among.
- ReAct — The perceive-act pattern it anchors.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome