Case note
Tool errors, handoffs, and when an AI agent should stop
A practical buyer question: when should an agent stop after a failed tool call or a missed handoff, instead of continuing to spend?
Agents do useful work by calling tools: tickets, code hosts, browsers, databases, internal APIs. Those calls fail. A handoff to a person is supposed to happen and does not. If the agent keeps going, it often repeats the failed call or drafts around missing evidence. The session stays active. The work does not.
Knowing when to stop is an operating rule. Some errors are transient and deserve a bounded number of repeats with backoff. Some errors are structural: the tool cannot do what was asked. Those should stop the agent and open a handoff, with the last error attached. Continuing past a structural error creates more usage lines, not progress.
Synthetic example. An agent is asked to update a record in an internal system. The tool returns a permission error. The agent asks the same tool again, then tries a related write, then drafts a summary that assumes the update happened. A reviewer later finds that nothing changed. Spend after the first permission error did not move the record. This illustration is not a customer story.
What to inspect
- Tool-call logs: tool name, success or error, error class if you have one, and timestamps.
- Sequences after an error: did the next step use new information, hand off, or repeat the failed call?
- Handoff records: when work was passed to a person, whether the last error went with it, and whether the item was completed afterward.
- Work records for abandoned or incomplete items, joined where possible to the same session or ticket identifier as the usage export.
If error class is not logged, you cannot write a precise stop rule yet. Start by logging the error string and the tool name.
For long-running work that may still be justified, see long-running work worth paying for. For cases where a non-agent tool already does the job, see When an existing tool beats an AI agent. Documented MCP tools live under docs/tools.
One next action
For a single workflow, list the last tool error before each abandoned session or human-taken session in the period your logs cover. Group those errors by class, then write one stop rule for the most common class: stop, hand off, or allow a bounded repeat. Attach the rule to the rows you used.
If you want the related spend and work exports in one place, 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 failed tool calls and missed handoffs can add spend without moving the work, and that stop and handoff rules should be written from logged error classes.
- It does not claim that agents should never repeat a failed call. It reports no error rates, spend totals, or completion rates, and it does not say how much any team would have avoided by stopping sooner.
- The short illustration is a synthetic example. It is not customer data and it is not a result.