Agent runs fail in the middle, not at the edges. Trace Explorer, now available to all private beta workspaces, gives you a step-level view of every run that passes through a Hyperpriors harness — and lets you replay any step with its original inputs pinned.
Highlights
- Step-level traces. Every model call, tool invocation, guardrail check, retry, and fallback appears as a discrete step with its full input, output, latency, and token cost.
- Deterministic replay. Select any step and re-execute it in isolation. Original prompts, tool results, and harness state are captured at run time, so replay reproduces what the model actually saw — not a reconstruction of it.
- Divergence view. Replay a step against a different model or prompt revision and diff the outputs side by side.
- Shareable trace links. Hand a colleague a URL to the exact step where a run went wrong, rather than a log excerpt and an apology.
- Retention controls. Trace payloads respect your workspace redaction rules; sensitive fields are masked before storage.
Why it matters
Debugging an agent by re-running it end to end is slow, expensive, and — because model outputs vary — frequently misleading. Capturing the full step context at run time makes failures reproducible artefacts rather than anecdotes. When a run misbehaves, you can now find the step, replay it, and fix the actual cause.