The West Coast Builders — coding-agent collection
Building the future of AI Coding Agents
Eight selected interviews on harnesses, multi-agent workflows, context, and review
Eight selected conversations with Jinjing Liang, Gil Feig, Ben Vinegar, Thorsten Ball, Li Yin, Quinn Slack, Mario Zechner, and Peter Steinberger from Mayank Gupta's broader interview playlist. Topics include agent harnesses, parallel worktrees, inspectable session history, review, architecture, model economics, and durable context. Product claims and forecasts remain the guests' accounts, not independently verified results.
In this collection Browse 8 summaries 5 of 8
SylphAI founder, AdaL creator, and former Meta AI engineer Li Yin joins host Mayank Gupta to discuss generated-code review, agent memory, human accountability, and her founder thesis for production-oriented coding agents.
Key Points Covered
- Generated-code volume creates a comprehension bottleneck: Yin argues that agents can produce tens of thousands of lines while human engineers remain responsible for understanding, reviewing, and maintaining the result; fast generation does not establish production quality 10:24–13:09(opens in a new tab).
- Review should begin with architecture: Her proposed AdaL workflow clusters large changes, reviews architecture before implementation details, preserves decisions across context windows, and proceeds to automated and manual testing 15:18–20:43(opens in a new tab).
- Coding and memory are separate responsibilities: Yin's “self-evolving” thesis keeps a code-writing agent focused on implementation while separate memory-bearing agents retain higher-level context and guide decisions 21:46–25:19(opens in a new tab).
- Humans remain accountable for consequences: She argues that automation can accelerate work but cannot accept responsibility to customers, so the human remains the final quality boundary 25:19–27:17(opens in a new tab).
- Visual review should complement local execution: Yin sees terminal agents as powerful because of their local shell access, but argues that humans need a more visual interface for understanding code and staying in the loop 29:18–33:10(opens in a new tab).
- Cloud sandboxes add testing capacity: She values cloud environments chiefly as parallel test environments, not as machinery for generating more pull requests that nobody has time to validate 34:03–35:56(opens in a new tab).
- Agent-era skill starts with defining the problem: Yin emphasizes asking the right questions, finding edge cases, explaining architecture, adapting across domains, and taking responsibility rather than treating junior or senior status alone as decisive 44:09–46:57(opens in a new tab).
- AdaL building AdaL is a first-party report: Yin says her team uses AdaL to build AdaL and selects models for different tasks, including UI and architecture work. This supports her product thesis but is not a demonstrated comparative outcome 46:57–48:52(opens in a new tab).
Full video: https://www.youtube.com/watch?v=5H862RhMgOU(opens in a new tab)