Tools & retrieval
What is Observation?
An observation is the structured result returned to the model after a tool runs.
Observations are the agent’s evidence. They should be logged, typed, and inspectable — not silently dropped.
Why it matters
Observations are the agent's evidence. Every time a tool runs, it returns structured data that the model uses to make its next decision. If observations are dropped, mangled, or unstructured, the model operates on guesswork instead of facts.
Good observation design is as important as good tool design. The observation format determines how well the model can reason about what happened.
Observation hygiene
Observations should be typed, timestamped, and logged. They should contain the tool name, the parameters used, the result data, and any errors. This makes the trace complete and debuggable.
Never silently drop a failed tool call. Record the error as an observation so the model can decide how to handle it.
Key takeaways
- 1Observations are the agent's evidence — they must be structured and complete.
- 2Never silently drop failed tool calls; log the error as an observation.
- 3Observation format directly affects model reasoning quality.
Common mistakes
- ✕Returning tool results as unstructured prose.
- ✕Dropping failed tool calls from the context.
- ✕Not including the tool name and parameters in the observation.
Related terms
Concept neighborhood
Terms linked from Observation in the glossary graph.
- Observation
- Tool Calling
- Trace
- Grounding