AI Engineer16 min
Give Your On-Call Its Own AI Teammate - Swaapnika Motapothula, Chinmay Gandhi & Vlad Tsvang, Meta
Swaapnika Motapothula, Chinmay Gandhi, Vlad Tsvang, Swapnika
Conference Presentation, Keynote
- Approximately 75% of new code is currently written with AI, with AI-generated changes reaching production daily despite the risk that high-quality code still causes outages requiring manual intervention.
- The team intends to deploy a dedicated AI teammate per rotation rather than adding automation incrementally, operating under the assumption that AI can both write and operate code.
- Agent autonomy will be granted via a graduated ladder: Level 1 covers investigation and notification, Level 2 involves acting with approval where most drafts are filtered by review, and Level 3 permits independent resolution of lower criticality work while high criticality tasks remain with humans.
- Initial system reliability measures include a 9 a.m. daily digest, scheduled monitoring to detect and repair stuck runs to prevent silent stalls, and a policy to page humans only when the system cannot resolve issues autonomously.
- Individual performance improvement is supported by a "bring your own agent" model allowing users to import personal skills and runbooks, with the system learning from every shift to ensure later shifts perform better than earlier ones.
- The agent's progression relies on earning levels through rigorous grading of proposed mitigations and incorporating corrections as memory when items are identified as wrong, with a future goal of earning medium criticality autonomy based on evidence.
- Team capabilities are expected to expand by converting chat interactions into executable skills without code writing, while future autonomy levels will require demonstrated evidence.
- Early indicators suggest promising reductions in on-call load, though the team plans to measure the time saved through rigorous studies of returned on-call hours rather than simple usage counts before assigning specific metrics.