The AI Harness: The Layer Between the Model and the Real World
PriyankaEveryone has been arguing about which foundation model is smartest. It's the wrong argument. The real bottleneck isn't the model -- it's the harness.
The harness is the software layer that lets a model actually do things. Control a computer. Execute code. Browse the web. Remember context across sessions. Call tools. Hand off tasks to other agents. Get evaluated against a real environment. Without a harness, GPT-5 is very expensive autocomplete. With the right harness, a 70B open source model outperforms it on real tasks.
Foundation models are commoditizing. Every leader gets caught to within a few benchmark points within six months. What differentiates real products is the harness -- and that's where the next generation of AI infrastructure companies is being built.
What Is a Harness
The word is new. The pieces aren't. A harness is the tooling stack around a language model that turns it from a text generator into a functional agent. It includes:
- Tool use and function calling. How the model invokes APIs, runs code, controls software.
- Execution sandboxes. The isolated environment where the model actually does the work -- a browser, a shell, a full desktop VM.
- Memory. Long-term context that persists across sessions, projects, and users.
- Computer and OS control. Screen understanding, mouse and keyboard control, file system access.
- RL environments. Simulated worlds where models are trained to complete tasks and get rewarded for succeeding.
- Evaluation infrastructure. Benchmarks, verifiers, and monitoring for whether the agent is actually working in production.
Why Now
Three things happened in 2025-2026 that turned harnesses from a niche into a category.
MCP became a standard. Anthropic released the Model Context Protocol in late 2024 and by 2026 it's the de facto way models connect to tools. OpenAI adopted it. Google adopted it. Every serious agent framework speaks MCP now. This is the HTTP-of-agents moment -- once there's a shared protocol for how models talk to the outside world, the ecosystem above and below that protocol explodes.
Computer-use agents proved the pattern. Anthropic's Claude Computer Use, OpenAI's Operator, and Google's Project Mariner shipped in late 2024 through 2025 and demonstrated that a language model can drive a full desktop or browser end-to-end. The prototypes are rough but the direction is settled: agents are going to control real software, not just call APIs.
RL post-training went mainstream. Fine-tuning with reinforcement learning moved from frontier labs to every serious open-source team. Which means everyone suddenly needs RL environments to train agents against. Someone has to build those environments, host them, and provide the evaluation infrastructure. That someone is a harness company.
The Independent Players
Nous Research ($50M Series A April 2025 led by Paradigm, plus a follow-on later that year). The ideological open-source counterweight to closed AGI labs. Building Psyche (decentralized training), DisTrO (distributed optimizer that reduces inter-GPU communication by orders of magnitude), the Hermes agent line, and a growing set of RL environments and agent tooling. Not a pure-play harness company but the training-and-environments side of the harness stack for the open-source community. Their bet: the harness should be open, portable, and not owned by whoever ships the biggest model.
Prime Intellect ($130M Series A July 2026 -- Founders Fund, NVIDIA, Andrej Karpathy, John Schulman, Clem Delangue, Dylan Patel). The most-funded harness pure-play in the market. Products include the Environments Hub (an RL environments marketplace), Lab (a full-stack platform for agentic post-training), and INTELLECT-3 (their own 100B+ MoE trained on their own stack). This is what a fully commercial harness company looks like -- environments plus training plus evals plus compute, all sold together. If any independent player becomes a $10B+ company in this space, Prime Intellect is the current front-runner.
OpenCUA (xlang-ai). Not a company -- an academic project from the University of Hong Kong with contributors at Moonshot AI, Stanford, Waterloo, and CMU. NeurIPS 2025 Spotlight paper. Open foundations for computer-use agents: the reference implementations, datasets, and benchmarks everyone else is building on top of. Watch the founders of this project -- they'll be the founding team of the next serious commercial harness startup.
trycua ("cua"). Open-source infrastructure for computer-use agents. Sandboxes, SDKs, and benchmarks for training and evaluating agents that control full desktops across macOS, Linux, and Windows. About 19,500 GitHub stars as of mid-2026. This is the commercial companion to what OpenCUA started as research -- productized secure sandboxes for enterprises deploying agents.
TinyHumans / OpenHuman. The consumer angle on the harness. OpenHuman is a personal AI agent with roughly 1 billion tokens of memory that bundles access to 30+ model providers under a single subscription. 20K+ GitHub stars in days after launch. Also building tiny.place -- a social economy for AI agents, with identity, discovery, encrypted messaging, and on-chain commerce between agents. The bet: consumers don't want to choose a model, they want one agent that lives across all their apps and remembers everything.
Others in the mix. Abundant.ai (RL environments and verifiers). Vecna Labs (open trajectory gym). Roder (open-source coding harness). LangChain and LangGraph (early framework, now arguably legacy but still deployed widely). Letta / MemGPT (the memory layer). Vercel AI SDK (developer harness for web apps). SupaEval (agent quality monitoring). Each of these is a slice of what a full harness needs.
The Big Labs
The big three foundation model labs all build harnesses -- but they build them for their own models and try to make the harness disappear inside the product.
Anthropic. Claude Computer Use plus MCP. MCP is the biggest development in the space and Anthropic gave it away as an open standard. Strategic move: if everyone speaks MCP, and MCP works best with Claude, Anthropic captures the harness value indirectly. The MCP bet is that being the protocol layer is more valuable than being one implementation of a harness.
OpenAI. Operator (computer-use), Codex CLI, ChatGPT Agent, and the Agent SDK. Playing a bundled game -- if you use OpenAI models, the OpenAI harness works best, so stay in the ecosystem.
Google. Gemini CLI, Project Mariner (browser agent), Agent Development Kit (ADK). Same play as OpenAI: proprietary harness tied to their model, with just enough interoperability to keep enterprise buyers happy.
The pattern is clear. The big labs are trying to absorb the harness into their model product. The independents are betting the harness should be model-agnostic and outlive any particular model.
Is It Investible? Three Ways
1. RL environments and training infrastructure (Prime Intellect model). Sell picks and shovels to everyone doing agentic post-training. Highest-quality revenue, defensible through data and specialized compute, potentially the biggest single outcome in the category. Prime Intellect's $130M Series A is proof point one. Expect at least two more well-funded competitors in the next 18 months.
2. Consumer / prosumer harness (OpenHuman model). Bundle the fragmented model market -- 30+ providers -- into one subscription with persistent memory. Sell as the personal AI operating system. Consumer subscription revenue, huge distribution potential, network effects if agent-to-agent commerce takes off.
3. Enterprise agent infrastructure (trycua model). Sell secure sandboxes, evaluation tools, and monitoring to enterprises deploying agents in production. Enterprise SaaS revenue, sticky contracts, clear buyer (VP Engineering or Chief AI Officer at every Fortune 1000 by 2027).
Business Models Across the Category
- Hosted training and compute. Prime Intellect Lab. Sell training runs and compute for post-training.
- Environments marketplace. Prime Intellect Environments Hub, Abundant. Long-tail marketplace of RL environments with a rev share.
- Consumer subscription bundling. OpenHuman. Pay one subscription, get memory plus all major model providers.
- Enterprise sandbox and eval SaaS. trycua, SupaEval. Sell to enterprises deploying agents.
- Open source with hosted commercial layer. Nous, LangChain. Community adoption for free, monetize on hosted deployment and enterprise support.
- Standards and ecosystem plays. Anthropic with MCP. Not directly monetized but strategically enormous.
The Contrarian Take
The harness is more valuable than the model. This is the argument that will look obvious in 2028 and controversial today.
Foundation models are commoditizing at the top end and getting smaller at the bottom end. The gap between the best closed model and the best open model shrinks every quarter. Meanwhile the harness -- the ability to remember, execute, control, verify, and coordinate multiple agents -- is where every real product actually spends its engineering effort.
Prime Intellect at $130M Series A is the market realizing this. Expect the pace of harness-layer funding to accelerate sharply through 2027, and expect at least one $10B+ outcome from an independent harness company by 2029.
What's Underpriced
Personal and consumer harnesses. OpenHuman is early. The consumer version of MCP -- one agent that lives across every app and remembers everything -- is a category with room for several winners and probably one dominant consumer brand.
Verticalized RL environments. Finance agents need financial modeling environments. Medical agents need clinical simulation environments. Legal agents need document workflows and case law verifiers. Someone is going to build the "Kaggle for agent training environments" for each vertical. Almost no funded competitors yet.
Agent-to-agent commerce infrastructure. tiny.place is early here. Once agents transact with each other -- pay each other, verify each other's identity, discover each other's capabilities -- someone has to run the identity, payments, and discovery layer. This is a category waiting for the right founding team.
Memory and long-context infrastructure. Letta, MemGPT, and their successors. Memory is the piece every harness needs and nobody has solved. The company that turns persistent multi-modal memory into a commodity API becomes core infrastructure for every agent stack.
Evaluation and observability. SupaEval, Braintrust, and others are early. As agents move from demos to production, "does the agent actually work?" becomes the dominant enterprise question. The Datadog-for-agents company is still up for grabs.
The Question Worth Debating
Does the harness collapse into the foundation model provider (Anthropic plus MCP becomes the default and squeezes everyone else into narrow verticals), or does it stay independent (Prime Intellect, Nous, OpenHuman and successors own the layer above the model)?
If the harness stays independent, this is a $50B+ category by 2030 with several IPO-scale winners. If it gets absorbed into the big three labs, the independent players get pushed into vertical niches -- still real businesses, just smaller.
The tell will be MCP adoption over the next 12 months. If OpenAI and Google continue supporting MCP as a genuine standard, the harness layer stays open. If any of the big three defect and push a proprietary alternative, expect the category to fragment along model provider lines.
Either way, the argument about which model is smartest is going to look quaint. The next generation of AI infrastructure is being built one layer up.