Skip to main content

​
KAHN Scope: per-agent dashboard ships

​
Headline

/#/agents is no longer a row-per-run table. It’s a row-per-agent dashboard — busiest first, outcome histogram per row, p50/p95 duration, mean convergence, sparkline trend, last-seen relative timestamp. The operator’s eye lands on the right thing in one render.

​
What changed

The Phase A surface answered “what runs ran?” The Phase B dashboard answers questions Phase A couldn’t:
  • Which agent is busiest? Rows ordered by run_count descending.
  • Which agent is failing? Outcome histogram (5-segment stacked bar, scaled across rows so cross-row comparison is honest).
  • How fast are agents converging? p50/p95 duration columns; linear-interpolation percentiles, byte-equivalent across FS+Postgres backends.
  • Is convergence trending up? Sparkline column with fixed [0,1] y-axis (cross-row comparable), endpoint dot coloured by the convergence-badge bucket of the latest score.
  • When was this agent last active? Relative-time last_seen.
Click any agent_id and the surface filters to that agent’s run list — Phase A’s row-per-run table is preserved as the inner panel. Both surfaces are reachable; neither has regressed.

​
The window selector

24h | 7d | all toggles the rollup window. Unknown values 400 with a documented allow-list — operators cannot typo new windows into existence.

​
Why this matters for pilots

The kahn.host external-pilot conversation through Phase A was “we can show you a list of CI runs with agent labels.” Phase B converts it into “we can show you which of your agents is currently degrading, before your customers tell you.” That’s the kahn.tools pitch made concrete.

​
What’s next

The dashboard is the substrate for two follow-on surfaces: cross-run audit flakiness (/#/audits, also shipped Phase B) and a future per-agent reliability heatmap (north-star Phase C deliverable). The aggregation endpoint serving the dashboard — /api/agent-runs/aggregate — is wire-shape-pinned in kahn-hq/docs/spec/agent-runs-aggregate.md and ready for SDK consumption. The landing-page rewrite at kahn.tools can now anchor on real screenshots of the dashboard, not placeholder copy. Specifically: a screenshot of the audit-runner row with its 6-event rollup (3 audits × pass+fail) is in tree as a check-in fixture and renders deterministically.