We keep watching people try to build the fully-autonomous agent. Point it at a goal, walk away, come back to finished work. We've tried it too, across about twenty agents running real parts of our company. It mostly doesn't happen, and the reason turns out to be more useful than the disappointment.
The first time we handed an agent a real multi-step job, research a list of prospects, draft the outreach, queue it, report back, it got three steps in and stopped. Not crashed. It hit a spot where it needed a decision we hadn't anticipated, made a reasonable-looking guess, and quietly went sideways. By the time we checked, it had produced a lot of confident, wrong work.
Autonomous is the wrong target
The mental model people reach for is a machine: set it up once, it runs forever. A workflow with twenty-five steps and real ambiguity in the middle almost never survives that. Every step is a place where the agent needs context it doesn't have, or a judgment call about what you'd actually want, and it only takes one wrong guess early to poison everything after it.
The model that works is the one you already know: a new direct report. Someone capable, fast, and genuinely useful, who still needs you to sync with them. You don't hand a new hire a quarter-long project and vanish. You check in, you unblock them, you answer the one question that saves them a day of wrong work, you share the context that was never written down anywhere.
The rhythm is the leverage
Here's the part that surprised us. The more you check in early, the more self-sufficient the agent gets later. Those first syncs are where you catch the wrong assumption before it compounds, and where the agent picks up the context it was missing. Skip them and you get exactly what you feared: a pile of plausible work you have to throw away.
So the leverage isn't in finding the magic prompt that lets an agent run untouched for a week. It's in the working rhythm. Short, frequent check-ins. Answer the blocking question. Hand over the context. Correct the small wrong turn before it becomes a big one. This is just human-agent collaboration, and it looks a lot like managing a person, because the failure modes are the same: missing context, an assumption nobody corrected, a decision made in the dark.
Do that for a stretch and the agent needs you less, because now it knows the things it kept getting wrong.
If you're building with agents
Stop optimizing for zero-touch. It's the wrong goal, and chasing it is why so many agent projects feel like they almost work and then don't.
Set the agent up, then treat the first stretch like onboarding a smart new hire. Watch closely. Sync often. Every time it does something wrong, ask why, and turn the answer into context it keeps for next time. The agents we trust with real work today are the ones we managed closely at the start, until they earned the longer leash.
The set-and-forget agent is a fantasy. The well-managed one is already doing real work.