AI agent observability, from trace to investigation

AI agent observability connects an outcome to the events that produced it. Start with a trace for the run, add events around the steps you need to investigate, and attach stable agent, workflow target, and external trace identifiers. These identifiers let an operator move from a reported incident to the relevant execution record.

The trace explorer supports filters and saved views. Failure groups bring related errors together, while recorded duration and cost fields help you investigate operational impact. Instrumentation determines what is visible: a model call or tool action needs a reported event before it can appear in the timeline.

Imagine a support agent that reads a ticket, requests account data, and then fails before drafting a reply. Locate the run using its external trace ID. Inspect the last successful event and the failure details, then use the agent and target filters to find other affected runs. A failure group can reveal whether the problem is isolated or recurring.

If the trace has the inputs and target context needed for replay, submit a variant with the configuration you want to compare. Robots Center records the new execution and its output differences. Use a controlled target when checking a fix: replay runs configured HTTP or connector actions and can cause real side effects.

An AI agent running on a robot produces software execution events while the robot reports its physical operating state. Send the agent's task and tool events as traces, and report the robot's heartbeats and telemetry through its own credential. Link the robot to its service-agent identity so operators can locate both records in the workspace.

For example, investigate a robot that stopped receiving work by checking its last heartbeat alongside the coordinating agent's reported task events. A low battery reading and a failed tool call suggest different next steps. This visibility depends on the device and agent runtimes sending the relevant data; the hardware's own control and safety systems remain responsible for physical behavior.

A grouped failure gives the team a shared place to investigate. Related traces and replay results provide evidence for deciding whether to change the runtime, adjust an integration, or add a regression case. Review recorded data with the workspace's redaction and retention settings in mind.

Start with one important workflow and a small set of diagnostic events. Give every run a stable correlation ID, record the outcome, and verify that the timeline answers a real incident question. Expand instrumentation where it helps operators distinguish failures, rather than collecting payloads without a purpose.