Agent Trace

Simple Definition

An agent trace is the step-by-step record of one AI agent run: what was asked, which models were called, which tools were used, what changed, what was retried, and how it ended.

Plain-English Explanation

Tracing is an established idea in software monitoring: following one request as it passes through different services. An agent trace applies the same idea to an AI agent. A typical trace might show the user request, a planning call to a model, a web search, a browser action, an API call, a test run that failed, a retry, and the final answer, each with timing, inputs, outputs, and cost.

The value is that it turns a black box into a timeline. Instead of trying to infer what the agent did from its final message, you can see what it actually did, in order.

Example

An agent answers a customer question with an outdated refund policy. The trace shows it searched the knowledge base, retrieved two documents, picked the older one because it ranked higher, and skipped the step that checks document dates. That points directly at the fix.

What Traces Are Used For

Debugging failed or strange runs, auditing which actions were taken and under whose approval, investigating security incidents, finding slow or expensive steps, building evaluation sets from real runs, and responding when something goes wrong in production.

Logs are raw events written by a system, often scattered across services. A trace organizes the relevant events into one connected execution path. Agent runtime observability is the larger practice of collecting and analyzing traces, logs, metrics, costs, and failures across many runs. A trace can contain sensitive data from prompts and tool results, so it needs the same access controls and retention rules as the data itself.