Instacart has open-sourced Blueberry, a Slack-native incident triage agent. In April 2026, it processed 25,000 diagnostic passes across 270 channels, delivering root-cause hypotheses in an average of three minutes per alert. The company built it to solve the real problem in the early minutes of any incident: not diagnosis, but coordination. Teams spend those minutes asking whether a deploy caused it, how broad the impact is, and what's actually happening.
Blueberry operates inside the Slack thread where the team is already working. When an alert fires — auto-triggered page, @mention, or active-thread follow-up — the system launches ten subagents in parallel. One targets internal knowledge: postmortems, runbooks, prior incidents. The others collect live signals simultaneously: deploy history, feature flag state, service ownership, error logs, anomaly sweeps. Wall-clock time is bounded by the slowest collector, not the sum. The result lands in the thread before most teams finish assembling.
The architecture has three explicit layers. Ingress is selective: Blueberry monitors Slack thread events and filters before queuing. The job management layer is a Postgres-backed durable queue, not in-memory. If a worker crashes mid-pass during a rolling deployment, another picks up from persisted state. The processor layer resolves channel ownership, loads a team profile, and mounts the right MCP endpoints before the agent accesses tools. Three MCP surfaces are stitched together: an in-process MCP for fast, stateful local helpers; a standalone shared MCP server for deploy analysis and dependency lookup; and remote, team-hosted MCPs that let individual teams expose proprietary data sources without touching the core.
Accuracy was the design forcing function. Blueberry's diagnostic accuracy climbed from the mid-60% range to the high 90% range after grounding it with 14 years of Instacart incident history. The company reported 99.9% workflow success rate, 58,000 MCP tool dispatches in a single month, and roughly 60 team profiles. One concrete example: Blueberry auto-triaged E2E test failures in staging and identified the root cause — a code change deployed before its database migration, triggering an ORM error on an undeclared enum attribute — before the incident channel formed.
| Metric | Value | Notes |
|---|---|---|
| Diagnostic accuracy — before grounding | mid-60% | Without 14-year incident history corpus |
| Diagnostic accuracy — after grounding | high 90% | After grounding on 14 years of Instacart incidents |
| Workflow success rate | 99.9% | Month-over-month across all passes |
| Diagnostic passes (April 2026) | 25,000 | Across 270 Slack channels |
| Slack channels monitored | 270 | April 2026 |
| MCP tool dispatches | 58,000 | In a single month |
| Team profiles loaded | ~60 | At time of open-source release |
| Average time to root-cause hypothesis | 3 min | Per alert, wall-clock |
Authority is deliberately constrained. Blueberry generates hypotheses and surfaces evidence; engineers retain every remediation decision. No automated production changes. CTO Anirban Kundu and VP of Engineering Siby Alappatt framed the system as a force multiplier for human judgment, not a replacement. Each diagnostic pass logs its reasoning trace, creating both a learning loop and an audit record.
For architects evaluating adoption, the hard part is the corpus. Fourteen years of structured incident history is what moved accuracy from mid-60s to high 90s. Teams without that corpus will start lower and need deliberate investment in incident documentation before grounding takes effect. The three-tier MCP design is the right abstraction: teams mount internal data sources without forking the core and scale heavy investigators independently. The Postgres-backed queue is the correct choice for any on-call system that must stay alive through its own infrastructure failures.
Blueberry is evaluable now. The open-source release lets you map integration cost against your Slack and observability stack this week, but budget time to build the incident history corpus before expecting high-90s accuracy.