Reports

Reports in the sidebar opens Agent Audit Reports: “Analyse a simulation run, an evaluation run or a trace: what the agent did, how it graded, and the specific issues behind the score.” A report is not a document you write or a file AgentHub stores. It is a view derived from data your project already has, recomputed every time you open it. There are three kinds of subject: Every simulation run, saved evaluation run and trace in the project automatically has a report. The only thing a report ever saves is a simulation run’s Suggested Fixes, which are stored on that run once generated.

​

The reports list

Under the heading, a line summarizes what’s available, for example “Derived from this project’s own data and recomputed each time a report is opened — 4 simulation runs · 2 evaluation runs · 120 traces.” Click a row or its name to open the report. Filters The list loads 20 at a time; Load more appends the next page. How status is bucketed: Empty states: Nothing to report on yet (with Go to Cost Calculator) when the project has no subjects, and Nothing matches these filters (with Clear filters) when filters hide everything.

​

Generate Report

+ Generate Report opens a dialog: “Pick what to analyse — a simulation run, an evaluation run, or a single trace — and the report is built from that subject’s own data.”

1. Report on: Simulation, Evaluation or Trace, with counts. Kinds with none in the project are disabled.
2. … to analyse: pick the subject. Drafts that never ran aren’t listed.
3. Click + Generate Report. The report opens.

The dialog tells you what generating costs: You can also reach reports directly: Report in the Traces list, Run report on a trace page.

​

What a report contains

Every report has the same layout: a header, a headline rate, a facts panel, a bar chart of results, Top Issues, and an evidence section. Sections that don’t apply to a subject say so and explain why, often with a link to the page that would fill them.

​

Header

The kind (“Report · Simulation run”, “Report · Evaluation run”, “Report · Trace”), the report name, a status, the subject ID, the date, extra tags such as environment and source, and the PDF and Markdown export buttons.

​

Simulation run reports

​

Evaluation run reports

​

Trace reports

Trace reports state plainly what they leave out: there is no scenario pass rate, no grader section unless an evaluation was run against the trace, and no suggested fixes for a trace on its own.

​

Suggested fixes (simulation runs)

On a simulation run report, Generate suggested fixes asks the run’s configured model to diagnose the failing scenarios and propose targeted fixes. If nothing failed, the button reads Suggest prompt improvements and the model reads the passing transcripts and grader averages to draft improvements instead.

• It is one real call, billed to the run’s model on your key. It can take a minute; keep the page open.
• The result is saved to the run and shown with “Generated … by <model> and saved to this run.” The list then shows Fixes generated.
• Each fix card has a Copy button.

It can’t run when:

​

Export and sharing

Both exports include what is on screen, including generated fixes. There is no public share link. To share a report inside your organization, send its URL: anyone with access to the same project in AgentHub can open it, and it is recomputed from current data when they do. To share outside, export it.