Skip to content
New here? Rent a GPU by the second
Kracht
Sign inSign up
Use cases
Studioplanned
Studioplanned

See what an agent run actually did, span by span

Every tool call, prompt and token recorded against the step that made it, so a bad answer traces back to the exact call responsible.

Everything below describes what this is being built to do. It needs Studio, which you can look around today but cannot yet use for the whole of this job. The problems are real now; the answers are the part still being built.

A wrong answer is easy to see. Which call produced it is not.

Why it is hard

  • A pipeline is a black box from the outside

    The input and the final answer are visible; every tool call, retry and token in between usually is not.

  • Logs are not a trace

    A pile of log lines from different steps has to be reassembled by hand into the order the run actually happened in.

  • The expensive step is a guess

    Without a per-step record, the step spending the most time or money is whichever one somebody suspects that week.

What answers it

  • Every call recorded against its step

    Spans, tool calls and tokens are attached to the node that made them, on the same graph the pipeline is drawn as.

  • The order is the recorded order

    A run replays as the sequence it actually was, not as a reconstruction from timestamps across separate logs.

  • The costly step is named, not guessed

    Time and tokens per step make the expensive one visible directly, which is the part not built yet.

story 13 of 20Next: Score a prompt change before you ship it

What you use from Kracht

Studio

Draw the workflow and attach hardware to each step, then read a finished run back step by step.

Related jobs

The graph exists to draw on now. The recording of a run against it is next.

Other use cases