Descriptions:
Brandon Walsenuk of Unblocked presents a practitioner’s guide to eliminating the “babysitting” problem in AI agent deployments—the constant need to manually supply context to agents that start every session with zero organizational knowledge. Drawing from lessons learned building Unblocked’s production context engine, the talk debunks three common myths (including the belief that larger context windows solve the problem) and delivers three hard-won architectural principles.
The core argument is that static context stores—CLAUDE.md files, curated document repositories—are necessary but insufficient. A production context engine must dynamically pull from distributed corporate knowledge sources at runtime (Slack, GitHub, issue trackers, internal wikis) and deliver a token-optimized, permission-respecting context packet to the agent before it begins work. Walsenuk argues that a social graph is the key indexing primitive: knowing which codebases an engineer works in, their PR history, and their organizational relationships lets the engine resolve conflicting information—for example, arbitrating between what source code says and what a CTO wrote in Slack—and surface what is actually authoritative.
The talk concludes with a preview of an open-source repository built at a workshop the prior day, containing a self-contained context engine component that teams can lift and adapt. Walsenuk’s company serves customers with up to 20,000 engineers, making permission-scoped context delivery via MCP with OAuth pass-through a hard production requirement. The session is aimed squarely at engineering teams looking to move from manual agent supervision toward autonomous, context-aware deployments.
📺 Source: AI Engineer · Published May 26, 2026
🏷️ Format: Deep Dive
