Last month, Tessera agents completed 2.6 million runs. We looked at every one that needed a human to step in, and the patterns were clearer than we expected.

Most failures start with the brief

Nearly half of escalations traced back to an ambiguous instruction — “clean up the CRM” rather than “merge duplicate contacts created this quarter.” Agents that asked a clarifying question first escalated far less often.

Memory is a reliability feature

Runs with access to workspace memory were markedly less likely to repeat a mistake a teammate had already corrected. Remembering is not a nice-to-have; it is how agents get better week over week.