Every workshop I run eventually gets to the same question: what happens if you just let it run. Nobody watching, nobody approving each step, the agent working through a task end to end on its own. It’s the question everyone actually wants answered, and it’s the one most demos are carefully built to avoid.
nanogent.ai exists specifically to answer it against real workflows, not a curated example.
The three failure modes that actually show up
Compounding small errors. A single wrong assumption early in a task doesn’t usually cause an immediate crash. It quietly shapes every step after it, so the failure you see at the end looks unrelated to the actual cause, which is three steps back. This is the hardest failure mode to debug, because the symptom and the cause are far apart.
Confident dead ends. An agent that hits a genuine blocker doesn’t always recognise it as one. It can keep producing plausible-looking output while making no real progress, because nothing in its process tells it to stop and flag uncertainty. A human doing the same task would feel stuck. An agent, left unattended, often doesn’t notice.
Scope creep on ambiguous instructions. Give a task a slightly underspecified goal and a person will usually ask a clarifying question. An unattended agent will often just pick an interpretation and run with it, sometimes the wrong one, and you won’t know until the output lands.
What to build before you remove the human
A checkpoint the agent has to pass before moving to the next stage, not just at the very end. A way for it to flag “I’m not confident about this” rather than always producing an answer. And logging detailed enough that when something does go wrong, you can trace it back to the actual decision point, not just the final broken output.
The honest takeaway
None of this means unattended agents don’t work. It means they work inside boundaries someone designed on purpose, which is exactly the workflow engineering argument this whole site is built on. The agent isn’t doing the directing. You still are, you’re just doing it upfront instead of step by step.