Instacart Builds Blueberry AI to Speed Up Incident Response

Instacart has introduced Blueberry, an AI-assisted incident response system designed to help on-call engineers troubleshoot production issues. When an alert triggers, the system launches approximately 10 subagents in parallel to analyze the problem. It pulls data from internal sources, including logs, deployments, service ownership, and other debugging signals, to generate a root cause hypothesis directly within a Slack thread. This initial investigation is typically delivered to engineers in about three minutes.

The system relies heavily on grounding its AI models with historical data. By incorporating more than 14 years of incident history, Instacart improved Blueberry's diagnostic accuracy from the mid-60% range to the high 90% range. During a single month in April, the assistant executed approximately 25,000 diagnostic passes across more than 270 Slack channels. Instacart reported a 99.9% workflow success rate, more than 58,000 Model Context Protocol tool dispatches, and support for around 60 team profiles.

For on-call engineers, Blueberry changes the starting point of an investigation. Instead of spending the critical early stages of an outage manually gathering logs, checking recent deployments, and identifying service owners, practitioners receive a pre-assembled package of context and hypotheses. Siby Alappatt, vice president of engineering at Instacart, called the system a "force multiplier in harnessing AI to help transform on-call" operations.

Importantly, Blueberry does not make automated changes to production environments. It uses a tool-aware harness that maintains the state of the investigation while leaving actual mitigation and remediation decisions entirely to human engineers. This ensures that while the AI accelerates the gathering of operational knowledge, human operators retain control over the actual resolution of system failures.