The fashionable version of long-running agents is “let it think longer.” The useful version is more grounded: every meaningful loop should leave behind a company-owned learning asset.
That asset can be small. A corrected rule. A failing eval seed. A checklist. A smoke test. A runbook paragraph. A script that verifies the next run instead of asking the model to remember what happened. Small is fine. Evaporating into chat history is not.
This matters because models are rented capacity. The compounding value is the operating memory around them: the domain rules, edge cases, verifiers, and workflows that stay useful when the model changes. If a loop does not improve one of those, it was probably just motion with a nice transcript.
Rule added: Every meaningful agent loop should improve a reusable local asset.