Are Harnesses Dead?

Summary

Theo explains what an AI harness is and argues that choosing one may become less important as models and agent tools evolve.

What an AI harness does

An LLM generates text; a harness lets it take actions. It gives a model an understood way to request operations such as reading or editing files, running commands, using memory, or calling other tools. Theo compares this to a chef needing an equipped kitchen: capability alone is not enough without tools to act.

An agent is an LLM paired with a harness. A harness may provide tools, context, system prompts, and an agent loop. Some, like Pi, are relatively minimal; others include file operations, shell access, skills, or subagents. Plugins can add further tools.

Interfaces, harnesses, and MCP

An interface is what a person uses; a harness supplies the model’s action capabilities. Theo characterizes the Claude Code CLI as an interface and the Claude Agent SDK as a way to use the underlying harness. Some products bundle both, making the distinction less obvious.

Theo presents MCP as a way for agents to connect to external systems, rather than necessarily as a harness tool. For example, MCP integrations can let an agent access databases or analytics through an agreed interface.

T3 Code

T3 Code is not a harness. It uses existing harnesses and adds a T3 Code MCP connection that can broker work across them—for example, by searching history, messaging other threads, or starting subagents with different harnesses. This lets users keep using a harness’s own skills and plugins while adding T3 Code’s cross-harness capabilities.

Theo describes T3 Code as an agentic development environment (ADE): an interface for both the user and the harness, with features for managing work and remote environments.

Making harnesses interchangeable is technically difficult because they have different behaviors and subagent systems. Theo says T3 Code takes on that complexity so users can switch harnesses more easily. He presents T3 Code as free and open source, while acknowledging that ADEs themselves may eventually become less central.

Why harnesses may matter less over time

Theo compares debates about harnesses to earlier arguments over tabs, editors, languages, and frameworks. As abstractions improve, developers can spend less effort on lower-level choices and focus on higher-level work.

He expects models and harnesses to become more closely integrated, with users choosing based on capability and cost rather than loyalty to a particular harness. He notes that differences between harnesses remain, but argues that many perceived advantages are small or likely to diminish as models improve.

Sponsored product: Eve

In the sponsored segment, Theo introduces Eve, a Vercel framework for building agents. Its project structure can define instructions, models, skills, TypeScript tools, sandboxes, communication channels, MCP connections, subagents, and scheduled runs. A Discord bot example shows an agent receiving instructions and making changes to a real codebase.

Guides and resources

The video is a conceptual explainer rather than a step-by-step tutorial or product review. Theo mentions a written “What is a harness?” article for people who prefer a reference to share with coworkers. Anthropic’s documentation is cited to distinguish the Claude Code CLI from the Agent SDK.

Speakers and sources

• Theo — sole speaker and creator of T3 Code.
• Vercel / Eve — sponsor and agent framework discussed in the sponsored segment.
• Anthropic documentation — referenced for Claude Code and the Agent SDK.
• Michael, creator of Effect — cited for a post expressing reduced concern about which harness is used.