Blueberry: Force Multiplier For The On-Call Engineer

This post details the architecture and design principles of Blueberry, Instacart's Slack-native on-call reasoning harness. It emphasizes the system's durability through a Postgres-backed durable queue for job management, its composition layer for resolving team-specific skills and MCP endpoints, and its three-tiered Model Context Protocol (MCP) surfaces (in-process, shared server, team-hosted) for flexible tool integration. The post highlights the agent's dynamic decision-making capabilities within a structured harness for reliability, and quantifies performance metrics like TTFI averaging 3 minutes and TTTT averaging 3 minutes for theory checks. It also provides a concrete example of SEV2 triage and an 8-turn PR exoneration scenario.