Cursor Cloud Agent v1 — conversation timeline rebuild

The v0 view was a flat log of API events. Helpful for engineers, useless for ops. I rebuilt it around a single question:“Can a non-technical operator see, in one screen, why the agent did what it did?”

• Prompts and thinking shown inline, not buried in JSON
• Tool calls expanded as cards with inputs, outputs, latency
• One reply per agent “turn” — threaded chronologically
• Failures highlighted with the exact line that triggered them

Cursor’s cloud-agent surfaces three separate streams — prompt deltas, thinking deltas, and tool-call events. They arrive interleaved, out of order, and at very different rates.

I designed the timeline as an append-only event logthat reconciles streams by causal links, not wall-clock. Every event carries its parent turn, so reorderings — or replays — produce the same coherent view.

Once every run is a reconciled timeline, evaluation gets easy. The judge harness reads the same event log a human would, scores the agent’s reasoning chain, and writes its verdict back into the same timeline.

Operators can now scroll to “why did this run fail the citation check?” and see the exact span the judge flagged. Debugging and grading became the same workflow.