Instacart built Blueberry, a Slack-native AI reasoning harness for on-call engineers that auto-triages alerts and supports multi-turn investigations. The system uses a Postgres-backed durable job queue, three composable Model Context Protocol surfaces (in-process, shared, and team-hosted), and parallel sub-agents to gather historical and live evidence simultaneously. In April 2026, it ran ~25k diagnostic passes across 270+ Slack channels, achieving ~3-minute average time to first insight and ~3-minute theory-testing turnaround with a 99.9% success rate. Key design lessons: harness quality (grounding, specialization, continuity, learning loops) matters more than model quality alone. Teams can contribute custom profiles, runbooks, and MCP endpoints to make tribal knowledge reusable infrastructure.
