Summary
Theo of t3.gg takes on a question dividing AI developers: do coding-agent harnesses still matter? Responding to the reaction to his own post comparing harness preferences to the tabs-versus-spaces debate, he lays out why he believes the harness is becoming a minor factor as models get stronger, while acknowledging the exceptions.
The video starts with a framework for how developers decide what to care about, from low-level style choices up to product and team decisions, and places harnesses within that stack. It then explains what a harness actually does, using a kitchen analogy: tools built into the harness are the equipment in the kitchen, while MCP servers are more like a cookbook the model can choose to consult. Theo explains why tools are tightly bound to a specific harness, whereas MCPs act as bindings to external systems, and uses the T3 Code MCP as an example of adding capabilities such as spinning up threads with other agents without interfering with Claude Code’s built-in behavior.
The sponsor segment introduces Vercel’s Eve, a framework for building agents from instruction files, skills, sandboxes, channels like Slack and Discord, and MCP connections. Developers building with Claude Code, Codex, or other agents will come away with a clearer mental model of where harness choice does and does not affect results.
📺 Source: Theo – t3․gg · Published October 30, 2026
🏷️ Format: Opinion Editorial
