Skip to main content

​
KAHN Phase B shipped — agent-fleet operator surfaces are live

​
What changed

KAHN Cloud now has four agent-fleet operator surfaces live behind the KAHN_AGENT_VIEWS flag. Phase A (2026-04-26 prod flip) made the agent ingest path exist. Phase B makes it useful.
SurfaceWhat it answers
/#/agents per-agent dashboardWhich agent is busiest? Which is failing? Is convergence trending?
/#/agents/<id> tool drilldownWhat did this agent’s failing tool actually do? (Click-to-expand input/output/error inline.)
/#/audits audit flakinessWhich audit checkpoint is currently degrading? (Deep-links to the failing run.)
/api/self/tenant per-tenant selfHow many agent runs do I have? Am I being rate-limited?

​
Why this matters for external pilot conversations

The kahn.host landing page through Phase A described a CI prototype. Phase B converts the live deployment into demo material that’s specifically about agent-fleet observability — the north-star pitch. A producer onboarding conversation can now show:
  1. Real per-agent rollup (not a row-per-run grep). Operators with 50+ agent runs/day can find the busiest agent, the most-flaky audit, the slowest tool, in three clicks.
  2. Convergence trends as first-class signal. The kahn-kiln north-star bullet (“convergence as observability”) now has a sparkline, a delta-vs-prior badge, and threshold-bucketed colour coding.
  3. Honest rate-limit visibility. Producers asking “is my emitter throttled?” see remaining tokens + a 5-minute throttled count without grepping logs.

​
Cost and risk profile

  • 10 gated tasksets, 0 rollbacks. 235/235 backend tests, 101/101 frontend tests.
  • Two substrate-class drift failures eliminated en route (retention parity for agent_runs; real migration runner replacing shell-history workflow). See companion learning on substrate-first sequencing.
  • One contract-extension PR avoided via inspection (D-probe gating on H1). See finding on D-probe H1 closure.

​
What’s next

Phase D (external pilot) is now unblocked. The frozen traceo-cat-pilot bundle has a realignment note and a probe-derived fixture set — a thaw is approximately a half-day of agent-shape rewrite, not a from-scratch onboarding. stratt-hq is the named candidate for a fresh probe; an A2A prompt template is filed for the next agent that opens that workspace pair. The landing-page rewrite (north-star Phase G G1) can now anchor on real screenshots of /#/agents instead of placeholder copy.