Case note
Agent loops vs long-running work worth paying for
A practical buyer question: how do you tell an agent that is stuck repeating itself from long-running work that is worth paying for?
Length alone is a poor judge. Some jobs should run long because the done definition is large: many files to review, a long checklist, a brief that has to cite sources. Those runs accrue time and billed usage for a reason you can point to. Other runs stay open because the agent covers the same ground or restates a plan without new information. They accrue time too, without moving the work.
The distinction is progress toward a written done definition. A short run can be a problem too, if it stopped before the definition was met and nobody recorded why. Ask of each stretch: did it add information, complete a step, or change the state of the work? Or did it repeat?
Synthetic example. Two agents each run for a long session. The first works through a numbered checklist and leaves new artifacts at each step. The second restates the same plan, queries the same store, and produces near-duplicate summaries until someone stops it. Both sessions look long on a duration chart. Only the first is justified by the work. This illustration is not customer data.
What to inspect
- Session and turn counts from orchestration or provider logs, including start time, end time, and stop reason if recorded.
- Traces that show whether successive steps introduced new inputs or outputs, or repeated the same call with the same arguments.
- Work records for the same period: ticket scope, checklist size, and whether the job was accepted as complete.
- Spend exports joined, where possible, to the same session or workflow identifier.
If you cannot see repeated steps in the trace, you cannot yet call a long run unjustified. Write the visibility gap down instead of inferring motive from duration.
When a long run ends in a handoff or a stop decision, see handoffs and when an agent should stop. After you change a run policy, measure it with Before and after: measuring AI quality, cost, and time. See also How it works.
One next action
Pull traces for the longest-running agent sessions in the last complete period your logs cover. Classify each session as progressing toward a written done definition, or repeating without new information. Keep the labels next to the session identifiers so someone else can check them.
If you want spend and work records for those sessions in one conversation, Overdrive helps you maximize your AI spend from inside your AI assistant. It is a remote MCP server with OAuth sign-in, and it works today in Claude, ChatGPT, and Grok Bot (setup steps). Start with a labeled sample pack and expand the lineage on each figure. Using Overdrive with your own company data is still in security review. The plan is for you to connect the tools you already use instead of uploading files, and those connectors are not available yet.
Claims and caveats
- This post claims that run length is not the same as value, and that justified long work shows progress toward a done definition while a repeated run does not.
- It does not claim that long runs are good or that short runs are cheap. It reports no duration cutoffs, spend totals, or shares of unjustified sessions, and it does not attach a money outcome to either pattern.
- The short illustration is a synthetic example. It is not customer data and it is not a result.