The decade of the agent
2025 was sold as the year of the agent. When it did not arrive, the date moved to 2026, and now to 2027. The direction is right. The timeline is fiction, and the fiction is expensive, because people are staffing and funding against it.
Here is the uncomfortable description of most systems currently shipped as agents: a language model, a tool-calling loop, and a prompt that asks it nicely to keep going. That is a real technique and it does real work. It is not autonomy. Autonomy would mean the system notices it is wrong, recovers, and finishes anyway. Almost nothing does that today.
Where they actually break
I run agent systems for my own work — task routing, tool access, automation around a local inference stack. They fail in a pattern that is boring and consistent.
The first failure is the error cascade. Step four gets a slightly wrong answer, step five treats it as ground truth, and by step eleven the system is confidently executing against a premise that was never real. Nothing crashed. There is no stack trace. The output looks finished.
The second is memory. Not the vector-store kind — the operational kind. What did I already try? What did I already rule out? Which of these three files did I edit, and did that edit land? Long-horizon work needs an accounting system, and shoving history into a context window is not one. It is a filing cabinet with no index.
The third is verification. An agent can generate work far faster than it can check work. So the bottleneck moves to whoever confirms the output was correct, and that is usually a human, which means the loop was never closed to begin with.
None of these are prompt problems. You cannot phrase your way out of a system that has no ground truth to check itself against.
The evidence points the same way
Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027 — as a forecast, not a measurement, but it matches what people say off the record. The deployments that survive tend to be narrow, supervised, and unglamorous: one task, one domain, a human on the approve button.
Karpathy’s framing is the one I keep coming back to. This is the decade of the agent, not the year. That is not a bearish position. A decade is fast. It is only disappointing if you promised a quarter.
What I do instead
The systems I trust have three properties.
They are checkpointed. The work is broken into steps that produce inspectable artefacts, so when something goes wrong I can find the step where the world diverged instead of re-running the whole thing and hoping.
They fail loudly. If a step cannot be verified, it stops and says so. A system that guesses when it is uncertain is worse than one that halts, because the guess propagates and the halt does not.
They keep a human at the decision points that are expensive to reverse. Not on every step — that defeats the purpose — but on the ones where being wrong costs more than being slow.
That is less impressive in a demo. It is the difference between something that works on stage and something that runs while I sleep.
What this predicts
The value in the next few years does not come from replacing a job with an agent. It comes from taking a task that used to cost an hour and making it cost four minutes, a thousand times over, with a person still holding the specification and the verdict.
The people who will look prescient in 2030 are not the ones who bet on full autonomy early. They are the ones who built the boring infrastructure — evaluation, recovery, audit trails, permissions — that autonomy will eventually need in order to be trustworthy. That work is unglamorous, it is available now, and almost nobody is fighting for it.
I would rather build that.