Every agent shipswith the dashboard that watches it.
We promise monitoring dashboards with every agent we build. Here is one — running on the same real recorded runs the Playground replays. Every number is computed from the transcript, never entered by hand.
live, replayable
wall-clock, all runs
9 distinct tools
plan · read · decide
self-corrected, no crash
declined vs hallucinated
Wall-clock per recorded run — how long the loop actually took.
What each run actually produced — read from its final payload.
- SQL AnalystAnswered
Answered with real figures and the SQL it ran.
- Document IntelligenceAnswered · cited
Answered from the corpus with 2 citations.
- Document Intelligence · refusalRefused · honest
Refused — answer not in the corpus, so it declined rather than invent one.
- Research & ReconSourced brief
Wrote a brief citing 2 sources it actually read.
- Recruitment ScreeningShortlisted
Ranked a 5-person shortlist, with deal-breaker rejections explained.
Every event of one real run on its true time axis. The gaps are the model thinking between actions — tools run in milliseconds; reasoning is where the wall-clock goes.
task · Which product category earns the most revenue from paid orders, and who are our top 3 customers by spend?
Hover or focus a mark to inspect the event — 8 in this run.
01 — One source of truth
The same real runs
Runtime, tool calls, reasoning steps, retries and outcomes are read straight off the recorded transcripts the Playground replays — derived in code, not typed into a slide.
02 — No invented metrics
What isn’t measured isn’t shown
These agents run through the Claude CLI with no API key and no per-call billing, so no token or cost figure exists — so we show none. A dashboard that invents a “$/run” is lying to you.
03 — Refusal is a feature
The honest-refusal count
One run declined to answer because the corpus didn’t hold it. That is tracked as a win, not an error — an agent that refuses cleanly is worth more than one that confidently makes things up.