Agent observability
Also known as AI agent observability, LLM agent observability
Definition
Agent observability is the practice of collecting and connecting evidence about an AI agent's decisions, model calls, tool use, state, errors, and outcomes. It makes a multi-step agent run inspectable after the fact.
Follow the task across steps
An agent can call several models, inspect files, execute tests, and revise its plan before opening a pull request. Looking at one model request at a time loses the relationship between those events.
Attach a stable task or trace identifier to each step. Record the selected model, input and output usage, tool name, arguments needed for debugging, timing, errors, and the state transition that followed. Remove secrets and respect the data boundaries of the systems being observed.
An illustrative investigation
Suppose an agent opens a pull request that fails a test. Its model calls completed successfully, but a tool returned an outdated dependency lockfile. The trace shows the tool result and the decision that followed, which is more useful than labeling the final patch a model failure.
Connect traces to outcomes
Useful outcomes might include tests passed, review accepted, task completed, cost, and time to completion. Keep these dimensions separate before combining them in a summary.
More telemetry is not automatically better. Choose attributes that support real investigations, control retention, and make privacy boundaries clear. The purpose is to understand and improve the agent workflow, not to record every token without a question in mind.
How this relates to Weave
Weave's agent and token intelligence views help teams connect AI usage and cost with engineering work. For a complete agent investigation, preserve the task identifier, model route, tool events, retries, and final outcome. A provider log that contains only the prompt and response cannot explain the full workflow.
Explore Token intelligence