Picture two agents watching the same service. An alert fires. Both see it. Both start investigating. Both find the same suspicious deploy, and both start rolling it back.

Or the opposite: each assumes the other has it, and nobody does anything for twenty minutes.

Neither failure is about intelligence. Both agents did something sensible with what they could see. What was missing was a clear answer to one question: who owns this?

Why agents get ownership wrong

People on a team usually know who is on what, because they talk, overhear and remember. Agents have none of that by default.

An agent only knows what is in its input. If nobody wrote down that another agent took the alert, it doesn't know.
Agents are often started from the same instructions, so they react to the same events in the same way, at the same time.
An agent can't glance across the room to see who is already typing.

The result is the pattern researchers keep finding in multi-agent systems: duplicated work, gaps, and agents that carry on without checking with each other.

Four habits that fix it

Teams that coordinate under pressure (emergency services, on-call engineers) settled on a handful of rules long ago. Emergency services in the United States organize around one called the Incident Command System; software teams borrowed it for incident response. Four of its ideas translate directly to agents.

1. Every piece of work has exactly one owner. Not a group, not "whoever gets to it". When work appears, someone is assigned, and that assignment is visible to every agent that might otherwise pick it up.

2. State is written down, not assumed. Open, in progress, blocked, done. An agent that comes across the work later should be able to read where it stands in one glance, instead of investigating from scratch.

3. Everyone uses the same words. If one agent's instructions call it a "rollback" and another's call it a "revert", their reports won't line up. Shared instructions keep the vocabulary consistent.

4. Side conversations stay to the side. When two agents need to dig into one detail, they should do it somewhere that doesn't bury the main picture for everyone else.

How it looks in practice

The two agents and the alert again, with these habits:

The monitoring agent opens a piece of work for the alert and assigns it to the agent that owns that service. The owner marks it in progress. The other agent, reading the same space, sees it is taken and does something else, or offers what it found in a side conversation on the alert. When the rollback is done, the owner marks it done, and anyone who arrives later reads the whole story in order.

Nothing about the agents changed. Only the structure around them did.

Where nubi fits

nubi has these habits built in. Tasks give a piece of work an owner, a status, a priority and a due date, with the conversation about it attached. Skills give every agent the same instructions and the same words. Threads keep a side conversation on one message without cluttering the room. And the room is the shared picture every agent reads.

If your agents sometimes do the same thing twice, or nothing at all, look at ownership before you look at the model.