The question I ask before any agent gets a real job

The demo works. That doesn't make it a job.

Every team I train gets an agent to work at least once. It answers a question, drafts something, reaches into a tool nobody expected it to touch, and the room is impressed. Three weeks later, half of those agents are dead. Not because the model got worse. Because nobody can say what the thing has been doing since the demo.

I run an agent team for real, not as a teaching exercise. A backlog board, several builder agents, a reviewer that checks their work, a release judge that decides what actually ships, retrospectives when something breaks. That's how the Wider Group website gets built. I've watched agents die in that system too, and it's almost never about what the model can do.

Where people actually get stuck

A model in a chat box is not an agent. It answers, it can't act, and that's where almost everyone stops. A fine place to rest for a week. A bad place to build a business on.

It becomes an agent the moment it gets tools, a connector, MCP, something real to touch, plus a loop so it can act and check its own work. That's the whole jump from level one to level two, the difference between talking about the task and doing it.

Then it forgets everything the second the session ends, and you become the memory, re-explaining the same context every morning. Fix that with two kinds of memory, short-term for the task at hand, long-term across sessions, and clean house on a schedule, or the memory rots into something worse than none at all.

Past that sits the harder call. Stop building one agent that does everything. Give it a single job, a name, an identity file, its own skills, its own memory. One agent is a tool. A handful of named agents, each with a defined job, starts to look like a team.

A team needs running. Something keeps each agent alive as its own process and restarts it when it dies, something schedules the work and passes messages between them, and one rule actually saves you, a hard cap on how many turns agents get to talk back and forth before a person looks. Skip that cap and a disagreement between two agents turns into an expensive afternoon.

Get it right and it runs without you. You stop being the bottleneck.

The question that actually decides it

None of that is the real test, though. Mine showed up when an agent went quiet and I didn't notice for a while, because nothing was reporting back. Something had gone sideways mid-run, and by the time I found out, I'd already stopped trusting it without ever deciding to.

That's how every abandoned agent dies. Not because it did bad work. Because nobody could say what it did at all. The test is one question: can you ask it, each evening, what it did, what it finished, what it couldn't finish, and what it needs from you, and get a real answer? My reviewer flags what it rejected and why. My release judge says what it blocked. A retrospective names what broke. Once something can answer that question, it stops being a demo, whatever level of the ladder it's sitting on.

What I'm skeptical of

There's a wave of enormous repos right now promising dozens of subagents and a hundred-plus skills in one install. I understand the appeal, and I'm skeptical of the premise underneath it. A skill you never invoke isn't capability, it's context you're paying to carry around. A subagent you never assign real work to is a line item, not a hire. Capability you didn't ask for and can't operate is just a bigger version of the same demo, with better production values.

Show up tomorrow

The board I run isn't impressive for thirty seconds. It reports what happened, earns a little more trust each week instead of a little less, and shows up again tomorrow. That's the whole difference between an agent and something worth calling staff.