Instacart's Blueberry AI cuts incident diagnosis time by automating early-stage triage for on-call engineers

Instacart has unveiled Blueberry, an AI-driven incident response tool that equips on-call engineers with automated triage and diagnostic hypotheses when production problems arise. Rather than replacing human judgment, the system accelerates the early-stage investigation work—gathering logs, checking deployments, identifying service owners, and surfacing relevant past incidents—that typically consumes the first minutes of an outage response.

How Blueberry Works#

The system operates as a collection of coordinated AI agents that run within Slack-based incident workflows. When an alert fires, Blueberry launches roughly 10 parallel subagents to investigate the issue, delivering initial findings directly into the incident channel within about three minutes. Rather than offering generic advice, each agent taps into Instacart's internal knowledge base: service ownership records, deployment histories, operational logs, and a 14-year archive of past incidents.

The critical design choice is grounding. Instead of relying solely on general language models, Blueberry links its reasoning to verified operational sources and maintains state as engineers investigate. Agents fetch information from connected tools while engineers retain full responsibility for diagnosis, mitigation calls, and actual fixes—the system proposes, humans decide.

Real-World Impact#

In April alone, Blueberry completed approximately 25,000 diagnostic cycles across more than 270 Slack channels. The platform achieved a 99.9% workflow success rate and dispatched over 58,000 tool calls to fetch contextual data. Accuracy in root-cause hypotheses climbed dramatically—from the mid-60% range to the high 90s—once the team trained the system on their full incident history rather than relying on generic models.

Why Operational Context Matters#

The engine behind Blueberry's effectiveness lies in its access to organizational knowledge. Rather than starting investigations from scratch, on-call engineers now begin with a curated set of logs, recent deployments, and analogous past incidents already retrieved. This reframes the troubleshooting starting point, allowing teams to focus on deeper analysis rather than data collection. The system also codifies lessons learned during incidents, making organizational experience—often trapped in team memory—available to future responders.