<- all tokdocs

Agent Harness vs Model: What Claude Code's Docs Actually Say

By , founder of Mavric and Apso

View on TikTok ->

An agent harness is the layer around a language model that supplies tools, context and the loop that chains actions together. A TikTok from @jakevanclief argues that the model is rented, the harness handles the looping, and the durable asset is your written instructions. Anthropic's Claude Code documentation backs that split almost word for word, but the video lists GitHub Copilot as a model when GitHub documents it as a product, and its story about Claude failing to reach a website blames the model for what is usually an access problem.

What is an agent harness?

The Claude Code glossary defines an agentic harness as "the tools, context management, and execution environment that turn a language model into a capable coding agent." It then says: "Claude Code is the harness; Claude is the model inside it. The harness supplies file access, shell execution, permission gating, memory loading, and the loop that chains actions together."

The companion page, How Claude Code works, describes the agentic loop as three blended phases: "gather context, take action, and verify results."

The video draws this on a chalkboard as three parts: instructions (what and how), the model, and what it can reach, with "HARNESS (Claude Code)" written underneath. Its closing checklist asks for the outcome you want, how you would do it, what the agent can touch and where a person checks.

The docs describe the harness handling the loop for coding and command-line work. The video's line that "for most people, the agent problem is solved" is the speaker's own generalization.

Is GitHub Copilot a model?

No. The chalkboard row labelled "THE MODEL" lists Claude, ChatGPT and Copilot. Only Claude names a model family.

GitHub describes Copilot as "an AI assistant that helps you write, understand, and ship software." Its supported models page opens with "GitHub Copilot supports multiple AI models, each with different strengths," and on 2026-10-08 the picker included models from Anthropic, OpenAI, Google, xAI, Microsoft and Moonshot AI. ChatGPT is also an interface, with GPT models underneath.

Copilot belongs in the harness row next to Claude Code. Putting a harness in the model slot is the kind of category mix-up the video says costs companies money. Model lineups change within weeks, so check the vendor page for the current list.

Can Claude Code access websites?

Yes. The video tells a story from a client job in Dubai where Claude "couldn't reach the government website it needed," so the same brief was run in ChatGPT and worked.

Claude Code ships web access as a built-in tool. The tool table in How Claude Code works lists a Web row for "Search the web, fetch documentation, look up error messages," and the glossary's definition of a tool includes searching the web. One government site failing to load usually comes down to permission rules, a corporate proxy, managed settings, or the site refusing the request. It does not show that the model lacks web access.

The anecdote does support the video's broader point. If the same brief moved between two systems and still produced the result, the brief was the asset.

Does Gartner say 70% of AI agent projects will be abandoned?

Gartner's figure is narrower than that. The video says "Gartner's saying up to 70% of companies will walk away from the agents vendors built for them by 2028." Techstrong.ai reported it on 2026-09-30: "By 2028, up to 70% of enterprises will abandon agentic AI systems built through vendor-assisted models as maintenance expenses soar and internal teams struggle to modify the technology independently." The quote is attributed to Gartner Senior Director Analyst Mukul Saha.

The prediction covers systems built through vendor-assisted engagements, such as forward-deployed engineering, where a vendor embeds its own engineers at a customer. The video's phrasing stays inside that scope.

It is a different prediction from Gartner's June 25, 2025 press release, which says over 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. That release also introduced the term "agent washing" for rebranded chatbots and RPA.

What is Interpretable Context Methodology?

The video shows the GitHub repo RinDig/Interpretable-Context-Methodology. It is real, MIT licensed, has 21 commits and credits Jake Van Clief. Its README opens with "Folder structure as agent architecture" and describes replacing framework-level orchestration with numbered folders for stages and markdown files for prompts.

The linked paper, arXiv:2603.16021, is titled "Interpretable Context Methodology: Folder Structure as Agentic Architecture." It was submitted on 2026-03-17 with a second version the next day. It has a second author, David McDermott, whom the video does not mention. The abstract still calls the system the Model Workspace Protocol (MWP), which matches the repo's commit history: "Initial release: MWP framework with 3 workspaces" followed by "Rename MWP to ICM in CLAUDE.md." An arXiv posting is a preprint and has not been peer reviewed.

What has no public record

The video opens with "A company paid me $120,000 to teach their people that agents are just a naming convention." An on-screen banner adds Dubai and three weeks. No client is named and no source confirms the fee, location or duration, so treat all three as self-reported.

The line "everybody's out here building their own AI agent from scratch" also has no source. The Gartner research cited later in the video describes the opposite problem: enterprises buying agents from vendors and then abandoning them.

Bottom line

The harness-versus-model split is accurate and matches Anthropic's own definitions. Copilot and ChatGPT are products that sit in the harness slot. When an agent cannot reach a site, check permissions and network access before switching models.

Sources

Published October 7, 2026. Based on a TikTok by @jakevanclief. Full video notes (markdown)