Descriptions:
Brandon Waselnuk presents at AI Dev 26 in San Francisco on what he calls the “context engine” — the infrastructure layer that allows AI coding agents to operate with the situational awareness of a senior engineer rather than a clueless new hire. The core premise is that every agent spun up in a terminal or Claude session starts with zero institutional knowledge, and closing that gap at useful speed requires purpose-built retrieval architecture, not ad hoc documentation or MCP connections alone.
Waselnuk walks through four myths about building context engines and three hard-earned lessons from his team’s production deployments. He identifies two common failure modes: static curated repositories that rot as soon as they’re created and require dedicated maintenance, and MCP pipelines that trigger “satisfaction of search” — a bias borrowed from radiology where agents stop searching once they find a plausible first answer, missing the actual root cause. A proper context engine must unify Slack threads, Notion pages, pull requests, and code history into a single graph-traversable knowledge base that can resolve conflicts between stale and current data.
The architecture his company has built ingests from existing team workflows without requiring behavioral change, builds embeddings and relational graphs for multi-hop retrieval, and surfaces context through CLI, API, and MCP interfaces with governance and token-scoping controls. An enterprise customer with 115,000 repositories and 30,000 developers is cited as a proof point for the conflict-resolution requirements at scale. Two open-source tools implementing components of the context engine are available for teams to adopt independently.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Deep Dive
