All posts Services Contact
Client login Get started

Why your agent quietly gives up on the long task

AI Agents Engineering

Someone hands the agent a twelve-step task. It gets through nine, produces something plausible, and reports back having quietly done a fraction of the work. Nobody was notified. From the outside this looks like the model being incapable. It is almost always one of three ceilings.

The iteration ceiling

Every tool round trip costs an iteration. A task that needs twelve real steps plus three decisions plus one correction needs twenty iterations before you account for reading the output of each step. Set the ceiling from the shape of the work, not from a round number, and — this is the part people skip — give a small grace allowance so a task that finished at iteration 89 of 90 is not discarded.

The context ceiling

By step nine the history holds three large tool outputs nobody has summarised. The model is now reading mostly its own transcript and is measurably worse at the task than it was at step two. Compress or summarise between phases rather than letting the history grow until something forces it.

The wall-clock ceiling

A long-running task that dies when a supervisor times it out is worse than one that never started, because the side effects already happened. Anything longer than a couple of minutes belongs in a background job with a durable identifier, not inside a synchronous request loop.

Design the task to survive all three

  • Break it into resumable phases with each phase persisted, so resuming does not restart from the beginning.
  • Check in at phase boundaries — a running summary in the task state, not a growing transcript.
  • Fail loudly on a ceiling and report exactly which phase it reached and what it completed.
  • Separate planning from execution. The plan is cheap to redo; the side effects are not.

The honesty rule

An agent that hits a ceiling and says "I completed steps 1 through 9, here is where I stopped, here is what remains" has behaved correctly. One that returns a confident summary of work it silently skipped has failed at the only job that mattered — telling the truth about what it did.