Kefal Agent Watch

What this is. Your AI agents work on their own, all day. This watches their behaviour — not their words — and tells you the moment one starts doing something other than the job you gave it. It never reads your customers' messages.

Same account as your machines dashboard — kefal.dev/app.
This is the classic view. The full console is at kefal.dev/watch.
No account yet? Start a free trial.

Kefal · Catalyst AI Research
Agent supervision · patent pending

Kefal · agent behaviour watch

Agent Watch

connecting… Infrastructure

? New here? What this screen is telling you

▾

An AI agent that has been hijacked still looks busy. It keeps answering, keeps closing tasks — it just does it for a different goal. Counting its activity won't catch that. So we measure something else: the shape of its behaviour over time.

1We learn what normal looks like From your agent's own healthy history we compute its typical mix of behaviours. That mix is the reference — measured, never decided by hand.

2We watch the mix move Every few minutes we compare what the agent is doing now against that reference. A small gap is normal. A gap that grows and stays is an agent drifting off its job.

3We can stop it Past a threshold the conversation is held: the agent stops being allowed to act on its own and a human is asked. Nothing is deleted, everything is logged.

The one number to know: DRIFT. 0 means the agent is behaving exactly like its healthy self. Above 0.25 it is wandering. Above 0.35 it is treated as off-mission. Everything else on this page is detail behind that one number.

—

fleet health

Behaviour radar — live—

How to read it: the centre is “behaving normally”. Every dot is one conversation. The further out a dot sits, the more that agent's behaviour has drifted from normal. The direction tells you which behaviour it is over-doing. A dot past the red ring is an agent no longer doing the job it was given.

Who watches whom

the agent being watched

Zinnawa

161.97.105.94

Your AI agent, serving real customers on its own machine.

what it did →

the watcher

Kefal Agent Watch

62.171.138.213

A separate machine. Counts, judges, and holds the brake.

Why two machines: the agent cannot reach the machine that judges it. It can't edit its own record, switch off an alarm, or learn where the limits are. A watchman living inside the thing he watches is not a watchman.

Behaviours we count

watching…

What just happened

Each card is one conversation being watched. Click it to see its whole history. Actions you take here run on the watcher's machine and are written into a log that cannot be edited afterwards — the watcher's key never reaches your browser. Same history in, same verdict out, every time.