Agents: the word of the year, and what survives production
Long autonomous loops still fail. Three-step flows with verification have been running for months.
Agents was the word of 2024. Long autonomous loops still fail. Everyone shipped something. Most of it doesn't work in production.
Every extra step multiplies the chance of drifting. A two-step loop has a certain error rate. A five-step loop has that rate to the fifth power. Past the fifth, the success rate falls off a cliff. We've seen it break at step three, four or five depending on the complexity of each step.
The promise was agency: set the goal and let it run. The reality was that every extra step needed its own evaluation, its own retry logic, its own fallback. The complexity compounded.
What does work: breaking the job into steps with acceptance criteria a piece of code can check. If the step produces structured output that validates, the loop knows it succeeded. If it fails, it retries. That bounded loop is reliable.
When a step can't be verified automatically, a person approves it. We haven't found a shortcut. A human in the loop isn't a failure of the system. It's the design. Handoff points with clear acceptance criteria keep the whole thing working.
The best agent is still mostly human oversight. Automation where it's safe, human judgment where it matters.