# Give Your On-Call Its Own AI Teammate - Swaapnika Motapothula, Chinmay Gandhi & Vlad Tsvang, Meta

## Executive summary

The talk introduces a novel AI teammate designed for on-call incident response, addressing the operational bottleneck that has emerged as AI accelerates code writing. Instead of relying on siloed automation (e.g., separate bots for alerts, tickets, or deploys), the system proposes a single, context-aware agent that connects all signals—alerts, tickets, incident history, and code changes. This teammate provides a human with a hypothesis, supporting evidence, and a proposed action, shifting the engineer's role from deep investigation to critical decision-making while ensuring all critical changes pass through normal human review and Continuous Integration (CI).

## Key takeaways

- The Operational Bottleneck: The challenge has shifted from writing code to operating it in production. On-call duty is the primary operational load, requiring a system that can synthesize information from multiple sources (alerts, tickets, history) simultaneously.
- Integrated AI Teammate: The proposed solution is one AI teammate per rotation that acts as a single loop across all signals, maintaining shared context and identity across shifts, unlike current siloed automation tools.
- Autonomy Ladder: AI autonomy is structured in levels: Level 1 (Investigation only), Level 2 (Drafts fixes for human approval), and Level 3 (Clears lower criticality work autonomously). Autonomy must be earned by demonstrating that mistakes are cheap to detect and catch.
- Human-in-the-Loop Safety: Critical production decisions and code changes remain human-owned. The AI teammate provides the evidence and the draft fix, but the human must approve the action, ensuring that review remains the primary filter.

## Technical details

- AI Agent Architecture: The core teammate is a simple agent utilizing a model with tools in a loop, acting on an environment and reading feedback. The system integrates team-specific skills, runbooks, and knowledge bases, allowing teams to 'bring their own agent' while leveraging the core platform.
- Shared Memory and Identity: The teammate's identity and memory belong to the *team* (the rotation), not an individual. This ensures that expertise, past decisions, and corrections flow into the shared memory, improving performance over time (e.g., Shift 30 > Shift 1).
- Feedback Loop and Learning: The system is designed to learn continuously: any correction or failure (e.g., when a proposed fix was wrong) reopens the item and is absorbed into the team's memory, making feedback an active input to the system.
- Automation Loops: The system includes automated loops for triage, recovery from stuck runs, and shift handoffs. It proactively monitors for stuck or failed runs and attempts to repair broken schedules before paging a human.

## Practical implications

- On-call engineers shift their focus from performing deep, manual investigations to making high-leverage, critical decisions based on synthesized evidence.
- The system reduces the overall cognitive load by handling routine alerts, duplicates, and long-tail noisy alerts autonomously.
- The process formalizes knowledge transfer by making the team's collective expertise and past decisions accessible to any engineer on call, regardless of shift.
- The system improves incident response by providing immediate, actionable hypotheses and proposed fixes, minimizing the time spent gathering context.

## Topics

AI Agents, Incident Response, On-Call Operations, Build Automation, Continuous Integration, LLMs, Swaapnika Motapothula, Meta, Chinmay Gandhi, Meta, Vlad Tsvang, Meta

Source: https://www.youtube.com/watch?v=_NBx2NNOuA4
