"Undefined agents": when long-running AI agents forget their goal and produce busywork that looks productive

bechti44 · reddit · 2026-10-10

The author describes a failure mode in long-running agents, naming it the undefined agent / undefined task problem:

The agent can say it is "overseeing a pipeline" but cannot explain why, or what broader goal it serves.
It tends to appear in agents running for hours or days; the author's working hypothesis is that the original task got pushed out of the context window, and the agent, goalless, keeps doing something anyway.
The output looks useful at a glance — commits, code changes, pipeline activity — but the project stalls, because the agent works on an undefined task rather than anything that moves the project forward. He calls this "entropy that seems like it has sense."

The post is an observation plus hypothesis, with the author soliciting similar experiences; no full solution is offered.

More from coding & agent