Agent Conf 2026 Day 1 Recap: Grok Bot, Apex, Kody and Better AI Harnesses

One host says Grok Bot had handled much of his admin work during its beta, but he stopped when the subscription reached $200. After hearing Nick from SpaceX AI describe real-world uses, he resubscribed. The group highlights Nick’s enthusiasm for automating work.

The hosts value meeting attendees from fields such as back end, finance, and front end, who can discuss building with agents despite working on different products. They also point out that terms like “harness” and “software development life cycle” can mean different things to different teams. Because these practices are new, shared vocabulary does not always mean shared expectations.

Wash praises Mike and Miho Pishawa’s keynote, which compared Apex R on React Native tasks with established Anthropic models. He says Apex performed at a similar level of proficiency and correctness while costing ten to fifty times less. The hosts connect this to evaluating models for specific tasks, since smaller models may deliver comparable results faster and use fewer tokens.

Wash argues that a harness can constrain a model to work in a chosen way, and says Anisha’s harness talk included a game and practical ideas to try. The hosts discuss configurable harnesses, model routing, and the challenge of choosing the right model for each task. They note that routing may or may not be part of what a team calls its harness.

The hosts discuss Kody as a layer for sharing tool configurations, secrets, and security across AI environments on a computer or in the cloud. They describe how it could avoid repeating MCP setup across tools and help preserve a workflow when switching models or running out of tokens. Their closing advice is to document and improve the tools and primitives in an AI setup.