"Fifty agents working while you sleep." The promise is everywhere. I put it to the test on my own fleet, and it blends two realities that have nothing in common. The real issue, in fact, isn't making AI more autonomous: it's knowing whether it's still running.
Two engines, only one runs at night
The promise sells "intelligent agents, around the clock". In practice, there are two engines, and only one works without you. Agents that reason only act during a session: they need a pilot and a context kept in hand. What really runs 24/7 is deterministic: a scheduled task, a queue that publishes. No reasoning, just reliable, predictable, debuggable mechanics. My own fleet taught me this the hard way: I had overused agents and underused deterministic automation, even though the latter is more reliable for any predictable flow.
What I take from it — Keep agents for reasoning, wire the predictable as deterministic.
The real trap: silent death
One of my synthesis routines stayed dead for 46 days, with nothing to warn me. On paper it was "running"; in reality it hadn't produced anything for weeks. It's the failure mode I fear most in a fleet of automations left to itself. The problem isn't capability, it's observability: the only question that matters is did it run? A brilliant agent that fails silently is worse than no agent at all: it keeps up the illusion that the work is done.
What I take from it — The question that matters isn't "is it intelligent?" but "did it run?".
What to put in place
- Sort your flows. What is predictable → deterministic automation. What requires real judgment → an agent. Don't hand an agent what a scheduled task does better.
- Set up a watchman. A check that verifies each active routine actually ran within its window, and alerts you if not. It's the building block that kills silent death.
- Require an artifact that gets read. Each routine must produce something you actually read. If you never read it, delete it.
- Bound the autonomy. Along a chain of steps, place a regular checkpoint (a human or a deterministic check), because an agent's error compounds at every step without a safeguard.
- Set a ceiling. Two to five routines that serve an active goal, and stop installing more. The time saved shouldn't immediately go back into configuration.
What I take from it — Two to five routines that produce an artifact someone reads and report their own failures beat fifty that rust.
Reality or myth
Reality: delegating repetitive tasks and finding prepared work in the morning. Myth: the autonomous army that builds a company while you sleep. Between the two, a word dream sellers never use: maintenance.
Putting AI to work on your processes without falling for this mirage is exactly what I do at Orogen. Write to me, let's talk.
Sources
- The full story of the routine that was dead for 46 days: see the resource The AI that dreamed.