Engineering · Platform
Operate the fleet — performance, policy, and the pipes
Instrumentation coverage 88% (7 of 8 reporting · 1 shadow). Fix next: Codex CLI is your slowest agent at 6,800 ms p95, costs $78.00, and is ungoverned. 3 agents are ungoverned. Latency measures speed, not task success.
Instrumentation coverage
7 of 8 reporting · 1 shadow
Worst p95 latency
Codex CLI · fleet p95 3,620 ms
Governance
3 ungoverned — no policy binding
MCP servers
tool surfaces in use · 52,400 events
Tail latency by agent — p50 → p95 → p99
5,000 ms slow line · ranked by p95Each bar runs p50 to p99; the marker is p95. Bars past the slow line are the tails to fix.
1 of 5 instrumented agents cross the slow line. Codex CLI owns the worst tail at 6,800 ms — fix it first.
Fleet trajectory
6-week governance indexPolicy mode
AuditObserve-only. An enforcing policy would block 3 ungoverned agents from acting.
Index is rising — ~2 weeks to the 80 target at the recent slope.
Your surfaces
Is it working? — task success, not just speed & cost.
LLM responsiveness — tail latency, throughput, slowest responses.
Every agent as an actor — runtime, model, host, and owner.
What an enforcing policy would block — Audit → Assist → Enforce.
Trace a run end to end when something breaks — the record behind every event.
Stream telemetry to your stack and close the shadow-AI gap.
Do next
Derived from app/data/c16-fleet.ts: agents · stats · performance.agents · postureIndexTrend · weeksToTarget · shadowCount. Fix-next crosses performance.agents (worst p95) with agents (cost + governed).