The Harness Thesis: How Anthropic Proved the Model Doesn't Matter, Then Lost Control of What Does

Here is the most important thing happening in AI right now, and almost nobody is framing it correctly: the company that best understood that the harness matters more than the model just antagonised the project that was proving them right.

Anthropic spent ten weeks shipping the most impressive product blitz in AI history — Claude Code, agent teams, auto-memory, an App Store, Code Review, Academy, Xcode MCP integration — a relentless demonstration that what wraps the model is where the value accrues. Then they sent a cease-and-desist to Nico Steinberger, the developer who’d built an open-source platform generalising that exact insight to every domain expert on the planet. OpenAI hired him two weeks later. Jensen Huang name-dropped his project at GTC.

This is a story about platform strategy, but it’s also a story about a company that wrote the playbook and then set it on fire.

The Abstraction Ladder

To understand what happened, you need to see the pattern in how AI developer tooling has evolved over the past eighteen months. It’s a clean abstraction ladder, and each rung represents a bet on how much you trust the model to do its job without scaffolding.

Rung 1: SDK wrappers. Google’s ADK requires six or more abstraction layers. OpenAI’s agent framework uses four. These are the “we don’t really trust the model” architectures — elaborate scaffolding of guardrails, routing logic, state machines, and retry handlers. The developer writes most of the intelligence. The model fills in gaps.

Rung 2: Minimal SDKs. Anthropic’s Claude Agent SDK stripped this to two layers. A bold move. It said: the model is good enough that you don’t need to babysit it with a state machine. Just give it tools and a goal.

Rung 3: Codeless. Claude Code and Co-Work eliminated the SDK entirely. Zero code. You talk to the model; the model uses tools; things happen. The developer’s job shifted from writing agent logic to describing intent.

Rung 4: Configurable codeless. OpenClaw. JSON files and markdown documents. No code, but also no CLI fluency required. You configure agents the way you’d configure a Notion database — structured enough to be precise, accessible enough that a supply chain manager or a litigation paralegal can do it.

Each step up this ladder represents the same implicit argument: the model is good enough; stop wrapping it in bubble wrap. The Google approach assumes the model is a fancy API that needs adult supervision. The OpenClaw approach assumes the model is a collaborator that needs context.

The data supports the latter. Opus 4.6 actually regressed on SWE-bench — the traditional coding benchmark — but jumped on agentic benchmarks. It got worse at being a code-completion engine and better at being a colleague. Average turn duration doubled from 25 minutes to 45 minutes between October 2025 and January 2026. These aren’t quick lookups anymore. They’re work sessions.

Anthropic understood this before anyone else. They built for it. And then they did the one thing you absolutely should not do when you’re winning the platform game.

The Blitz

First, let’s give credit where it’s obviously due.

Between January and March 2026, Anthropic shipped ten major features in ten weeks. This is not normal. This is not even particularly sane. But it was effective.

And here’s the number that matters: companies running Claude Code at scale were processing 334 million tokens per day. Per company. One enterprise saw 80% PM adoption within six weeks of rollout. These aren’t pilot numbers. These are “the old way of working is dead” numbers.

The blitz proved something that a lot of model-obsessed commentary still hasn’t internalised: the harness beats the model. The UX beats the benchmark. The workflow beats the weights. Anthropic didn’t just believe this. They proved it, empirically, at scale, with revenue.

The Fumble

On January 27, 2026, Anthropic’s legal team sent a cease-and-desist letter to Nico Steinberger.

Steinberger had built Clawdbot — an open-source platform that let non-developers configure and deploy AI agents using JSON files and markdown. It sat on top of Claude’s API. It was, architecturally, the logical next step of Anthropic’s own abstraction thesis: if Claude Code proved you don’t need an SDK, Clawdbot proved you don’t need a CLI.

The C&D forced a rename. Clawdbot became Moltbot, then OpenClaw. Anthropic followed up with an OAuth lockdown restricting consumer API tokens — a move widely interpreted as targeting OpenClaw’s authentication flow. There was zero public endorsement. Zero acknowledgment. Zero “hey, cool, someone is building the platform layer we haven’t gotten to yet.”

Instead: lawyers.

Let’s pause on the strategic absurdity of this. Anthropic had just spent ten weeks proving that the harness matters more than the model. OpenClaw was building the harness layer that Anthropic hadn’t built yet — the one that puts agent configuration in the hands of domain experts, not developers. Every OpenClaw deployment ran on Claude’s API, generating revenue for Anthropic.

The usage data tells the story: OpenClaw traffic as a percentage of Claude Code volume went from 5.5% on January 26 to 28.9% by March 9. That’s not cannibalisation. That’s a flywheel. Claude Code’s biggest npm download spike coincided with OpenClaw’s ACP integration.

The Gift

On February 14 — Valentine’s Day, because reality has a sense of humour — OpenAI hired Steinberger.

Within weeks, Codex had full OpenClaw compatibility. OpenAI didn’t just tolerate the open platform; they ensured it worked seamlessly with their stack. The contrast was not subtle.

Then Jensen Huang, at GTC in March, said the words out loud: “OpenClaw sparked the agent era.”

By this point, OpenClaw had 321,000 GitHub stars, 1,239 contributors, and 2.32 million npm downloads per week. It had become the de facto standard for codeless agent configuration. And it was now, thanks to Anthropic’s legal team, conspicuously model-agnostic.

Here is what Anthropic handed OpenAI: not a developer, not a codebase, but a strategic narrative. The story went from “OpenClaw is the open-source extension of Claude’s ecosystem” to “OpenClaw is the model-agnostic platform that works with everyone — and Anthropic is the one company that tried to kill it.” In platform markets, narrative is adoption. Adoption is distribution. Distribution is everything.

The Two-Layer Problem

This brings us to the structural issue, which is more interesting than the corporate drama.

The harness strategy now has two distinct layers:

Layer 1: Proprietary harness. Claude Code, Codex, Gemini Code Assist. These are controlled, first-party experiences optimised for developers. They’re excellent. They’re also, by design, walled gardens. Claude Code works with Claude. Codex works with OpenAI’s models. The value capture is clean: you use the harness, you pay for the tokens.

Layer 2: Open configurable platform. OpenClaw. Model-agnostic, community-driven, designed for domain experts who will never open a terminal. JSON configs, markdown instructions, a plugin ecosystem. The value capture is… complicated.

Here’s why Layer 2 matters more for the next phase of AI:

The enterprise data is stark. 68% of companies are in exploration or pilot phases. Only 11% are in production. But the companies that reach production go all in — 334 million tokens per day, 80% team adoption, fundamentally restructured workflows. The gap between “piloting” and “production” isn’t a model quality problem. IDC data says only 6% of CIOs consider their organisations AI-ready. 50% of AI initiatives stall post-pilot. 89% cite data quality problems rooted in uncaptured domain knowledge.

The bottleneck isn’t the model. It isn’t even the harness, in the Claude Code sense. It’s the gap between what the model can do and what domain experts know how to ask it to do. The supply chain manager who knows that Tuesday deliveries from Supplier X are always 15% short — that knowledge isn’t in any database. It’s in her head. And she’s not going to learn Python to encode it into an agent.

OpenClaw solves this. JSON configs and markdown files are the interface between domain expertise and model capability. It’s the layer that turns tacit knowledge into agent behaviour.

And Anthropic doesn’t own it. Doesn’t influence it. Actively antagonised it.

The Domain Gradient

There’s a reason coding was the first domain where AI agents achieved real traction, and it isn’t just that AI companies are staffed by programmers (though that helps).

Coding is a formal domain. It has verification infrastructure — compilers, test suites, type checkers, CI pipelines. You can tell a coding agent to fix a bug and objectively determine whether it succeeded. The feedback loop is tight. The error signal is clear. Trust can be calibrated.

Now look at the trust numbers for AI agents generally: 3% of enterprises report high trust. Unsupervised delegation ranges from 0–20%. These are abysmal numbers. And they make perfect sense, because most domains aren’t formal.

Semi-formal and situated domains can’t rely on compilers to verify agent output. They need a different trust architecture — one built on domain expert oversight, iterative configuration, and transparent reasoning. They need, in other words, exactly what OpenClaw provides: a way for the person who knows the domain to configure, monitor, and adjust the agent without writing code.

The Anthropic Paradox

Let me state the paradox as cleanly as I can:

Anthropic built the intellectual framework for understanding that value migrates from model to harness. They executed a ten-week product blitz that proved it empirically. They shipped the best developer-facing AI harness on the market. And then, when a third-party project appeared that would extend their harness thesis to the 89% of enterprise knowledge that’s currently uncaptured — the exact unlock needed to move enterprises from pilot to production — they sent lawyers instead of engineers.

There are generous interpretations. Brand protection is real. “Claw” in the name was a trademark concern. The OAuth lockdown may have been about security, not suppression. Corporate legal departments have their own momentum, disconnected from product strategy.

But generous interpretations don’t change outcomes. The outcome is:

• OpenClaw is model-agnostic and growing at 2.32M npm downloads per week
• OpenAI has the creator on staff and full platform compatibility
• The open-platform narrative belongs to OpenAI, not Anthropic
• The enterprise domain-expert layer — the one that matters for the next $100B of AI revenue — is developing outside Anthropic’s ecosystem
• Jensen Huang is crediting OpenClaw, not Claude Code, as the spark for the agent era

Anthropic still has the best proprietary harness. Claude Code is excellent. The blitz created genuine switching costs. For developers, Anthropic is arguably still winning.

But the next hundred million AI users aren’t developers. They’re the domain experts who know things no model has been trained on. They’re the operations managers and compliance officers and logistics coordinators whose tacit knowledge is the actual bottleneck to AI adoption. And the platform being built to reach them just got pushed into a competitor’s arms.

What Happens Next

The harness thesis isn’t wrong. If anything, Anthropic underestimated how right it was.

Value is migrating from model to harness. The best model doesn’t win; the best experience does. But “harness” isn’t one thing — it’s a stack, and the stack has layers that one company can’t own. The proprietary layer (Claude Code) and the open platform layer (OpenClaw) were, briefly, complementary. They could have been symbiotic. Claude as the preferred model inside an open ecosystem, the way Postgres is the preferred database inside the cloud-native stack — not because it’s mandated, but because the experience is best.

Instead, the layers are now adversarial. OpenClaw will optimise for model-agnosticism. Developers will build agents that work with Claude and GPT and Gemini. The model becomes commodity. The configuration layer becomes the platform. And Anthropic, which understood all of this before anyone else, will watch the value migrate one layer above them.

The 68% of enterprises stuck in pilot mode are going to break through. When they do, it won’t be because the models got better. It’ll be because the interfaces got accessible enough that domain experts could encode their knowledge into agent configurations. That’s the unlock. That’s the next phase. And the company best positioned to own it decided, in January 2026, that a trademark mattered more than a platform.

History may record this as the moment Anthropic proved its own thesis — and then handed the conclusion to someone else.

---

The numbers in this piece are drawn from npm registry data, GitHub metrics, IDC enterprise surveys, and internal adoption reports as of March 2026.