Summary
Jeffrey Lee-Chan, an engineer at Snapchat, runs a hands-on workshop at AI Engineer on achieving “idea velocity” using Open Claw — an open-source framework that extends Claude Code with persistent task memory, contextual history, and multi-agent parallelization via git work trees. The session targets both newcomers setting up Claude Code for the first time and advanced users exploring experimental orchestration patterns.
Lee-Chan’s architecture separates concerns between two layers: Open Claw acts as the orchestrator, retaining goals, task specs, and project history (like his “skeptic agent” custom code review system), while Claude Code handles implementation. The rationale is that keeping orchestrator context focused on what to accomplish — rather than how — preserves decision quality as agent count scales. Parallelization is achieved through tmux terminals and git work trees, enabling multiple simultaneous Claude Code instances on isolated branches. The agent orchestrator framework he forked from open source manages worker coordination at the bottom of the stack.
The workshop also covers a practical testing taxonomy for agent-built features: automated tests for logic, browser tests for visual correctness, and the shrinking category of tasks that genuinely require human judgment — a category Lee-Chan notes has contracted significantly in just six months as agent reliability has improved. Two live demo projects illustrate the system in production: an AI RPG with D&D-style dice mechanics that generates reactive story worlds, and a multi-model research aggregator that queries several LLMs in parallel, combines outputs, and surfaces answers with side-by-side tab notifications.
📺 Source: AI Engineer · Published July 11, 2026
🏷️ Format: Tutorial Demo
