The Harness, Not the model

Last week, we talked about the diffusion curve of AI, the idea that adoption is still concentrated inside clusters of power users, and that the more interesting question isn't whether people are for or against AI, but how quickly it becomes embedded across different kinds of work. We argued we're barely out of the first inning, with the score still mostly off the board.

This week we want to look at what actually moves us along that curve. Because it's tempting to assume the answer is just "better models." A new frontier release drops, benchmarks tick up, and we treat that as the engine of diffusion. We'd argue it's increasingly something less glamorous and far more consequential: the harness around the model.

If you haven't run into the term, an agent harness is the scaffolding that turns a raw model into something that can actually do work. The model reasons, the harness gives it memory, tools, scheduling, file access, and a way to act without being prompted at every step. In other words, the model enables reasoning, but the harness enables capability. The reason a system can suddenly book your travel, triage your inbox, or refactor a codebase isn't that the underlying model got mystically smarter, it's that systems engineering amplified what the model could already execute.

This is the part of the stack that's been quietly exploding, and most of it is open source.

Two harnesses worth watching

The clearest example of the wave is OpenClaw, a fully open-source (MIT License) agent harness that bills itself as an "operating system" for AI agents. It manages model calls, memory, and channels so the agent doesn't have to. It's local-first, memory is stored as plain Markdown files on your own machine, and it runs a heartbeat daemon (a computer program that runs in the background) that lets the agent act on a schedule without waiting for you to type anything. The model is an interchangeable module: you can point it at Claude, GPT, Gemini, DeepSeek, or a local Llama through Ollama, and swap as you like. Out of the box, it connects to thousands of community skills such as: web search, email, calendar, file management, even smart-home control. It went viral because it makes a commodity model feel like a capable employee.

But the one I keep coming back to is Hermes Agent, released by Nous Research in early 2026. On the surface it looks similar to OpenClaw: open source, MIT licensed, local-first, no telemetry, installs with a single command, and works across hundreds of models through OpenRouter and direct provider integrations. The repo reportedly went from zero to over 57,000 GitHub stars in six weeks, faster than OpenClaw at the same stage. That alone tells you something about where developer energy is going.

What makes Hermes different, and why I think it matters more for the diffusion story, is what it does after it finishes a task.

Hermes and the self-improving loop

Most agents are amnesiacs. They complete a job, the context window clears, and the next time you ask for the same thing they start from scratch. Hermes is built around the opposite premise. It lives on your server, remembers what it learns, and gets more capable the longer it runs.

The mechanism is the interesting bit. After completing a task, Hermes automatically distills what it did into reusable "Skills" and stores them in persistent memory, a closed-loop learning system. But it goes further than personal memory. Under the hood, Hermes is also a platform for generating training data: it can spin up thousands of tool-calling trajectories in parallel, compress them to fit a token budget, and export the whole thing in ShareGPT format (the standard shape for fine-tuning datasets). It ships with native integration to Nous's own reinforcement-learning stack, so the experience an agent accumulates in the field can be fed straight back into training a better model.

The harness isn't just a way to use a model. It's a machine for producing the exact data needed to improve the next one.

Why this distills through the whole ecosystem

Here's where it ties back to last week. The diffusion of AI doesn't only happen when a non-technical person finally opens a chatbot. It also happens underneath the surface, in the supply chain of how these systems get built, and that's where Hermes is quietly doing something structural.

For most of the last few years, the data that made frontier models good at acting (calling tools, navigating multi-step tasks, recovering from errors) was scarce and locked inside a handful of labs. It's expensive to manufacture and even more expensive to label. A harness like Hermes flips that. Every developer running it on real-world tasks is, almost as a byproduct, generating high-quality agentic trajectories. There are already community projects built specifically to turn real tasks into fine-tuning data for the next generation of Hermes models, plus a skills hub designed as an open standard so that a skill learned in Hermes is portable to Claude Code, Cursor, Codex, and others.

That portability is the distillation we want you to notice. A capability discovered by one agent, in one company, doing one mundane task, doesn't stay there. It gets compressed into a skill, exported as a dataset, folded into an open standard, and absorbed by the rest of the ecosystem. The smart behavior of the few becomes the baseline of the many, not through a press release, but through shared infrastructure.

What this means for a business

If you run an organization, the practical takeaway is that the cost of getting AI to do real, specific work in your environment is collapsing, and it's collapsing in the open. You no longer need a frontier lab's permission or a six-figure platform contract to give a capable model memory, tools, and the ability to act on a schedule. You need a harness, a model you can swap out as prices fall, and a few real workflows to point it at.

The strategic shift is subtle. A year ago, the durable advantage looked like access to the best model. Increasingly it looks like the accumulated skills your agents have learned doing your specific work, your distilled trajectories, your private memory, your captured institutional knowledge. The model is becoming the interchangeable module. The harness, and what it remembers about your business, is becoming the asset.

So if last week's question was who isn't using AI yet, this week's is while the rest of the curve catches up, the tools to compound a real advantage are sitting in open-source repos right now, free, waiting to be pointed at a problem. The model gets the headlines. The harness quietly does the work and, increasingly, teaches the next model how to do it too.

We may still be early in the inning. But the players who figure out the harness aren't waiting for a better bat. They're already on base.