Meta’s on-call AI teammate earns autonomy one action at a time

Meta engineers Chinmay Gandhi, Swaapnika Motapothula and Vlad Tsvang describe an AI teammate for on-call rotations. Their aim is to connect alerts, tickets, deployments and incident history so engineers receive evidence and a decision to make.

In an illustrative 2 a.m. incident, twelve alerts arrive together. A conventional assistant can investigate when asked, but the engineer still chooses every next step and connects the findings. The proposed teammate starts triage earlier and prepares a hypothesis, evidence and a possible fix.

Separate alert, ticket, deployment and incident tools each hold part of the context. The team proposes one teammate per rotation, with its own identity, shared team memory and a human approval gate.

The autonomy ladder begins with investigation and suggestions only. At the next level the agent drafts a fix for human approval. A third level handles selected lower-criticality work while high-criticality production decisions remain with people.

An action earns more autonomy only when an error is inexpensive and easy to catch, and the agent has a track record on that particular action. Code changes still pass through normal review and continuous integration.

An illustrative board shows 23 incoming items, with three requiring a human decision, two in progress and 18 handled. The speakers use synthetic example data; they do not claim these figures as measured production savings.

Three automation loops triage incoming signals, look for stuck or failed runs, and prepare shift handoffs. The system can route an alert, task or test failure to an appropriate investigator and notify a person when it needs help.

Teams can supply their own agents, skills and runbooks. Runs are logged so another engineer can inspect the trajectory or resume the conversation instead of starting from scratch.

The teammate belongs to the rotation rather than an individual. Its team-owned identity has only the access the team grants, and shared memory carries useful decisions and corrections between shifts.

Memory must remain useful as circumstances change: entries can be scored, promoted or retired. When someone corrects an answer, the work can reopen and the correction becomes new evidence for the team.

The team says early signs of reduced on-call load are promising but reserves a numerical savings claim for a rigorous study. Next steps include improving accepted mitigations, incorporating feedback and measuring time saved.