Meta’s Chinmay Gandhi, Swaapnika Motapothula and Vlad Tsvang describe an AI teammate for an on-call rotation, not just a chatbot for one engineer. Their illustrated 2 a.m. response is a scenario, and they explicitly defer a measured time-savings claim.
Keep one context across the shift
- Connect alerts, deployment diffs, incident history, runbooks and tickets so the on-call person receives a hypothesis with evidence instead of stitching together separate bot replies.
- Give the teammate its own rotation-scoped identity and shared memory. Record decisions and corrections, attribute remembered advice and retire entries that stop helping.
- Use recurring automations for triage, stuck-run detection and shift handoff; offer a morning digest of what is handled, pending and awaiting a human decision.
Earn, rather than assume, autonomy
- Level one investigates without changing production. Level two drafts fixes that need approval. Level three handles selected low-criticality items whose mistakes are cheap and detectable.
- Critical mitigations stay human-owned. The team says code changes still go through normal review and CI; most drafts do not land.
- Trust should be specific to an action and its track record, not a global permission granted because the model sounds capable.
What remains unmeasured
- A demo board’s 23 incoming items and three human decisions illustrate the intended workflow, not a measured average across real rotations.
- The speakers describe early encouraging signs, but say a rigorous study of actual on-call time saved is future work. Memory promotion and automated recovery also need independent failure checks.
“Autonomy has to be earned.” — Chinmay Gandhi