One-Line Summary: As frontier models commoditize within a benchmark point of each other, the harness — not the model — is what users adopt, customize, get locked into, and pay for; the harness layer captures most of the durable value in the agent economy.
Prerequisites: What is an AI harness, the 2026 harness landscape
The Thesis
For a user shipping real work in 2026, the model is interchangeable and the harness is not. This was not the case in 2023, when GPT-4 was the only model that could reliably power an agent. By mid-2026 there are four or five labs producing models within a single benchmark point of each other on most agentic tasks, often at radically different prices, sometimes shipping new versions monthly. The model is increasingly a commodity input.
The harness is the opposite. Switching from Claude Code to Cursor means re-learning a UX, re-authoring CLAUDE.md files into .cursorrules, re-writing hooks (Cursor has none), re-installing plugins through a different marketplace, and losing every saved agent definition. That switching cost is the moat. It does not look like the moat the model labs assumed they were building, but it is the moat that matters.
What the Evidence Shows
Three patterns from 2024–2026 made the thesis hard to ignore:
- Frontier model swaps inside harnesses are routine. A Claude Code user moving from Sonnet 4.5 to Opus 4.7 keeps everything: hooks, plugins, saved memory, sub-agent definitions, slash commands. The harness is the unit of investment; the model is a parameter.
- Harness-vendor revenue grew faster than per-token API revenue. Cursor's ARR crossed $200M in 2024 with no model of their own. Ruflo runs ~100k MAU on a free tier with paid plans for federation. Anthropic and OpenAI started shipping first-party harnesses (Claude Code, Codex CLI) precisely because the harness is where retention lives.
- Multi-provider harnesses succeeded. Ruflo, Cursor, Continue, Aider all support multiple model providers. Users do not seem to care which model is behind the curtain when the harness experience is what they value.
Why It Matters for Course Choices
A direct consequence of the thesis is that the most durable skills are harness-level: knowing how to extend a harness, design a sub-agent, write a plugin, configure a topology, debug a hook. These transfer across model upgrades. Model-specific skills (prompt-engineering tricks tied to a specific Claude version) age fast.
A second consequence: the strategic question for any team building agents in production is not "which model" but "which harness, and what should our extensions look like."
Counterargument and Where It Fails
The naive counterargument: "But model intelligence is the most important variable in agent quality." The data suggests it was, and now it isn't — at least for tasks within current model capability. Two same-cohort frontier models drop in for each other inside the same harness without dramatic quality differences for most tasks. Where they do differ (long-horizon coding, mathematical reasoning, niche languages), users are not switching harnesses; they are switching models inside the harness.
Where the counterargument has teeth: tasks at the absolute frontier of model capability. If your agent is solving problems no current model has been seen to solve, the model is still the dominant variable. For everyone else — most users, most tasks — the harness is the product.
How Harnesses & Frameworks Implement This
| Question | Single-agent harness | Orchestration framework | Framework / SDK |
|---|---|---|---|
| User loyalty target | The harness brand (Cursor, Claude Code) | The platform brand (ruflo) | Your application brand |
| Switching cost | High (config + memory + hooks) | Very high (also: topology + workflows) | Owned by your application |
| Model-swap pain | Low | Low | Owned by your application |
Connections to Other Concepts
what-is-an-ai-harness.md— The category this concept argues is the product.the-2026-harness-landscape.md— The roster of brands competing for that loyalty.harness-cost-models.md— How harness-level decisions dominate cost.choosing-your-harness-stack.md— Capstone decision.
Further Reading
- Codex Blog, "Claude Multi-Agent Ecosystem" (2026) — Survey of how harness-level products consolidated.
- Anthropic, "Building Effective Agents" (2024) — Early articulation of the harness as the unit of engineering.