← Glossary
Glossary · AI and agent workflows
Observation
The result the harness returns into the context after executing a tool: contents, a listing, a status or an error.
An observation is evidence that an operation ran and of what it returned, and no more. A listing shows names, not contents; an error describes the failure the tool reported, while a timeout can leave a remote action's outcome unknown. A correct observation can still be misread by the model, so the output must be checked against the observations it was built from.