How do I know what my AI agents actually did?

Faith doesn't scale

One agent you can watch. A fleet you cannot. Running several agents with tool and commit access without seeing what they do is the fastest way to lose trust in the whole setup.

What to capture

Per agent and per task: what it did, what it cost, where it failed, and whether the result passed your bar — close to real time. That's the difference between operating on evidence and operating on hope (it's what yoru is for).

Catch the silent failures

The failures that hurt report success while doing nothing. Observability plus reconciliation against ground truth is how a stale cache or a frozen queue surfaces from a system, not from a client.

Straight answers.

Isn't logging enough?

Raw logs aren't observability. You need per-task cost, outcome, and failure visible and queryable in near real time, tied to whether the work met your bar.

Why does a fleet need this more than one agent?

Because you can't personally watch many agents at once. Observability is what lets a human operate a fleet instead of babysitting one assistant.

What does it catch that tests don't?

Silent, fail-green production issues — work that reports success while doing nothing. Tests check code; observability plus reconciliation checks reality.