Operate the fleet — and know it actually works.
For engineering and platform — the one seat that runs the agents and proves the runs.
Latency and cost charts tell you how fast and how expensive a model is. They never tell you whether the task succeeded. The control plane measures all three axes — speed, outcome, and behavior — on a fleet you can actually see.
Fast and cheap tells you nothing about whether it worked.
An agent can run fast, run cheap, and fail half its tasks — and every dashboard you own would call it healthy. The metrics that are easy to capture measure the engine, never the result.
The engine, not the outcome
Latency and spend get watched because they're easy. Whether the agent did its job goes unmeasured, so a retrying, failing agent looks efficient.
Averages hide the tail
A healthy mean sits on top of a long tail of slow, costly calls. The p95 and p99 that users actually feel never show up in a single number.
No way to trace the break
When something goes wrong, a latency chart can't tell you what the agent did. You're left guessing at a run you can't replay.
Three axes: is it fast, did it work, and what exactly did it do?
One seat that measures speed, outcome, and behavior on the same fleet — with runtime inventory, policy config, and the telemetry pipes underneath.
Speed — Performance
p50, p95, and p99 per agent, with cost per call alongside — so the slow-AND-expensive compounds rank to the top on their own, not a flattering average.
Outcome — Effectiveness
Task success rate, intervention rate, and rework per agent — the “is it working?” axis latency can't capture, each labeled measured or inferred.
Behavior — Activity Log
Trace a run end to end — tool calls, file access, MCP traffic, LLM turns in order — and follow it from start to the action that went wrong.
Inventory, policy, and pipes — the ground truth all three axes rest on.
Runtime inventory with lifecycle hygiene
Every agent and its runtime detail — model, binary, cwd, owner — with active, stale, and shadow states, so orphaned and dark agents surface instead of hiding.
Policy config you can rehearse
See exactly which agents an enforcing policy would block before you arm it — named agents, named actions, run in observe-only first.
Telemetry pipes to your stack
Stream events to the systems you already run, with the export gap named — the shadow agents that emit nothing are called out, not silently dropped.
Honest about the signal
Every outcome number is labeled measured, inferred, or none — so you know which metrics are evidence and which are still estimates. The shadow agents that leave no trail are counted as a gap, not zero.
The same fleet, every axis
Speed, outcome, and behavior read from one inventory. The slow agent, the failing agent, and the misbehaving run are the same actors — correlated, not three disconnected dashboards.
Fix the worst thing first
The seat names one prioritized target — worst tail latency crossed with cost and governance — so the next move is the highest-leverage one, not a scattered list.
Run the fleet on evidence, not averages.
Speed, outcome, and behavior — on a fleet you can see, trace, and fix.