I read agent papers every day.

This week I tested an agent that is already working in production, with a real role and a scoreboard.

I have been paying attention to where agent systems actually break. The breakage shows up in tool access, state, and handoff, the unglamorous layer nobody writes papers about.

So I gave @viktor_com a real job in my own workspace to see how it handles that problem.

It is an AI employee that lives in Slack. Three things stood out.

The integration layer is the moat. It reaches around 3,000 tools through one connection, scoped OAuth, SOC 2, rather than a graph of brittle per-tool auth.

It carries a standing brief instead of a prompt chain. Role, standards, sign-off rules, persistent across sessions. This is closer to onboarding than to prompting.

Human-in-the-loop is part of the design. Every action is proposed, a person approves, and the trail is there afterwards. That is what makes it trustworthy enough to leave running.

One I did not even ask for. It caught a metric being reported on two different windows across two connected sources, and proposed a fix.

Customer side, Antonin Stetina, CEO of KULINA Group: "Mindblowing all-in-one AI which does everything in a single solution." Five campaigns became sixty across 29 markets on $2.5M in spend, no headcount added.

The bigger change is who owns the work. Viktor runs the whole job end-to-end and hands you proposals to approve.

If you are building agents, look closely at how this handles orchestration and permissions. Happy to go deeper on that if it is useful.