Traces you can actually read

One question, asked a thousand times

When we interviewed thirty teams about how they debug agents, every conversation came back to the same sentence: “Why did it do that?” They were not asking about latency or token counts. They wanted to know what the agent saw, what it decided and which piece of context tipped the decision.

The old trace viewer could not answer that quickly. It showed every event in one long list, ordered by time. A typical run produced 140 rows. Finding the decision that mattered meant scrolling, expanding and copying JSON into a text editor. Median time to answer “why” in our usability sessions was just over eleven minutes.

What we changed

We rebuilt the viewer around decisions instead of events. Each step in a run now shows three things side by side.

• What it knew. The context the model had at that moment, with new information since the previous step highlighted.
• What it chose. The action, its arguments and the short reason the agent gave.
• What happened. The tool response, trimmed to the fields the next step actually used.

Raw events are still one click away. Most people never open them.

A trace is not a log. A log tells you what happened. A trace should tell you why.

Diffs between runs

The most requested feature was comparison. When an agent passes a case on Monday and fails it on Thursday, you want to see where the two runs split. The new viewer lines up two runs step by step and marks the first point where the context or the choice differs. In testing, that cut the time to find a regression from roughly twenty minutes to under three.

Approvals in the timeline

Human approvals used to live in a separate audit screen. Now they sit in the trace where they happened, with who approved, how long it took and what they saw. Teams with strict review rules told us this alone saves them a weekly export.

The numbers so far

The new viewer has been in beta with 120 teams for six weeks.

1. Median time to answer “why did it do that” fell from 11 minutes to 2.
2. Support tickets tagged “can’t understand trace” dropped by 70% among beta teams.
3. Shared trace links, our best signal that people trust what they see, more than doubled.

You can turn it on today from your workspace settings. It works on every existing run, with no migration, and becomes the default in the next major runtime release. Follow along on the changelog.