The West Coast Builders — coding-agent collection
Is Agent Harness relevant in 2026
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 3 of 8
Ben Vinegar argues that specialized agents still have a role because no model invocation can carry every company's knowledge. The Modem co-founder and former Sentry VP of Engineering joins Mayank Gupta to discuss task-specific memory, evaluations, and harnesses—the software that gives models tools and manages their work.
Key Points Covered
- Modem connects customer feedback to shipped work: Vinegar describes a cloud agent that remembers customer requests and product activity. After a pull request merges, it can identify affected customers and prompt the team to follow up. It can also respond to chat requests or work with a coding agent through MCP. These are his product claims, not independently demonstrated capabilities 39:15–43:34(opens in a new tab).
- Generated-code share depends on engineering practice: Vinegar reports that 99.9% of Modem's code is generated and argues that navigable code and isolated components help agents work effectively. The percentage concerns code generation, not productivity or quality, and the referenced experiments are not shown 47:21–50:35(opens in a new tab).
- Tool narratives reflect vendor incentives: He cautions that model vendors and influencers benefit from claims that general agents can replace software, while distinguishing low-stakes generated utilities from important tools his company prefers vendors to maintain 51:39–55:01(opens in a new tab).
- Finite context leaves room for specialization: Vinegar argues that agents cannot carry unlimited memory and that task-specific memory, skills, and evaluations can support vertical products. He says Modem spent thousands of dollars on evaluations, but supplies no results or comparison 56:38–58:31(opens in a new tab).
- Harness design remains open to experimentation: He points to continuing changes in coding-agent interfaces as evidence that harnesses are not settled, and characterizes Modem as a cloud harness that borrows coding-agent concepts without being a coding agent 59:20–1:04:10(opens in a new tab).
- Expertise requires staying with a problem: Vinegar advises builders to spend enough time in a problem area to develop distinctive knowledge, separating thematically related experiments from repeatedly chasing unrelated quick wins 1:04:59–1:06:39(opens in a new tab).
Full video: https://www.youtube.com/watch?v=wbsDZpVB4Lg(opens in a new tab)