On September 5, the error rate for a tool used by three of our GTM agents started climbing. It went from 1.7% of calls failing to 14.8% over the next week, roughly 1 in 7 calls instead of 1 in 60.
The agents use this tool to write structured records into our team workspace, such as adding a contact, updating a status, or moving a record. Their tasks were still getting done, so we didn’t notice the increase by looking at completed tasks.
How we found it
We run Tessary on our own agents. Its tool error classifier uses CUSUM to track changes in failure rate, so it can catch both sudden spikes and slower degradation over time. When this tool crossed the threshold, it opened a finding showing when the increase started, the before and after rates, and examples of failed calls.
Tessary’s triage and root cause analysis investigated the finding and traced the failures to four mistakes:
- An agent wrote a value that wasn’t a valid option for a select field.
- A write call was missing a required identifier or used an undocumented action argument.
- An agent tried to move a record by writing a link into its body. That changed the content, but didn’t move the record.
- An agent attempted a change that could have deleted content beneath a record, and the store refused it.
The three agents don’t share code. They do share this tool and its instructions. Tessary identified those shared definitions as the place to address the mistakes and suggested the documentation change.
What we changed
We improved the shared tool documentation that all three agents use. The tool definitions now explain how to check select values against the live schema, which identifiers and actions a write requires, how to move a record, and when a deletion warning needs human review.
We made the change once, in the instructions for the tool, so every agent using it gets the same guidance.
If you have multiple agents writing through a shared tool, it’s worth tracking the tool’s error rate alongside task success. In our case, the tasks alone wouldn’t have shown how often the agents were making calls the store refused.