AI Harnesses Are the Software Shells That Turn AI Models Into Capable Agents

Set Trending Topics as a preferred source on Google.

When people talk about artificial intelligence, they usually mean the models. Names like Claude, GPT and Gemini stand in for the capabilities of entire AI systems. For agents, that view falls short. A language model can produce text or propose a tool call. To edit files on its own, run research, execute programs and pursue a goal across many steps, it needs additional software.

That operating layer is increasingly called an agent harness. A binding definition has yet to emerge. Microsoft describes it as the runtime scaffolding that turns a language model into a working agent: model and tool calls, conversation state, context, and rules for approvals and multi-step tasks. The research paper Harness-Bench defines it as the mechanism that organizes context, tools, state, permissions, constraints and error recovery between the model’s output and the outside world. In shorthand: agent equals model plus harness.

What an AI Harness Does

A harness leaves a model’s trained weights untouched. It creates the conditions under which the model can act, essentially through four functions.

First, instructions and policies. A system prompt defines the agent’s role, how it should prioritize, and when it must stop or ask for approval. These rules work like a job description handed to the model at runtime as binding context.

Second, tools. Web search, a terminal, a file system, an email client or interfaces to enterprise software. The harness tells the model what a tool is for, executes the call the model selects, and returns the result. Open standards such as MCP and Agent Skills, developed by Anthropic and later released as an open standard, are meant to make those capabilities portable.

Third, the agentic loop. An agent interprets the task, picks an action, examines the result and decides whether another step is needed. The harness keeps that loop running while capping runtime, cost and the number of iterations. Microsoft’s agent harness supports automatic context compaction, approval rules and bounded loops for exactly that purpose.

Fourth, context and memory. Long tasks eventually stop fitting into a model’s context window. The harness summarizes older information, offloads data to files or databases, and retrieves it again when needed. Logging, permissions and isolated execution environments come on top. All of this determines whether an agent still remembers its original goal after a hundred steps.

A translation layer can also make several models available behind the same interface. Model-agnostic harnesses such as OpenCode, Pi and Hermes Agent plug into multiple providers or open models. Systems like Claude Code, Codex and Gemini CLI perform similar functions while staying tied to their respective AI ecosystems.

Why the Software Layer Is Gaining Importance

Several studies now show that the same model core delivers markedly different results depending on the harness around it. In one study of harness effects, Claude Opus 4.5 scored 45.9 percent on SWE-bench Pro inside a standardized scaffold and 55.4 percent with Claude Code. Across six leading models in the standardized scaffold, the spread was only 4.9 percentage points.

Costs vary widely too, a problem many teams already feel in daily use, for instance when agents top up their own credit automatically. Databricks compared identical models with the same compute budget across different coding harnesses. In some cases the cost per task differed by more than a factor of two while output quality stayed the same. One reason the company cites: the minimalist Pi harness sent roughly three times less context per step.

The harness became especially visible on ARC-AGI-3, an interactive benchmark built from unfamiliar video games. Nvidia reported that its research architecture Agentic Variation Operators, or AVO, solved all 183 levels of the public set using Claude Opus 5. It combines a custom execution loop, persistent memory and a supervisor that detects dead ends and redirects the main agent.

The headline often drawn from this, 30 percent without a harness and 100 percent with one, oversimplifies. The ARC Prize reported 30.2 percent for Claude Opus 5 in its own test, yet the two numbers come from different experimental setups. A controlled comparison that swapped only the harness is still missing. Scale AI’s HarnessOpt-Bench counsels similar caution: when optimizing harnesses, the choice of optimizer model had on average roughly 1.8 times the effect of the choice of harness. What matters, then, is how the two components work together.

From Technical Tool to Strategic Layer

For companies, the harness could become valuable because it holds an agent’s durable configuration. Models are relatively easy to swap through an API. Much harder to move are workflows, permissions, tools, accumulated experience, review routines and the history of completed tasks. That shifts where lock-in sits: an open harness reduces dependence on a single model provider, while attachment to the harness platform grows once an agent’s institutional knowledge lives there. For users, that makes open data formats, exportable sessions, portable skills and the option to run the harness locally or with another operator just as relevant as multi-model support.

Local systems promise more control over data and infrastructure. OpenClaw, which turned a weekend project into an open-source hit, positions itself as a personal assistant running on the user’s own devices and connects the agent to channels like Slack, Signal, Telegram and WhatsApp. Hermes Agent can be self-hosted and run with various models and local, Docker, SSH or cloud sandboxes. Whether that adds up to full data sovereignty depends on which model and which external services get connected afterwards.

Which Business Models Are Emerging

Several overlapping revenue models are taking shape around harnesses.

Open source with commercial add-ons: the core stays freely available while the vendor earns from hosting, team features or governance. Under Earendil’s licensing model for Pi, the core remains MIT-licensed while additional products can be offered as fair source, proprietary software or a cloud service. OpenCode pairs an open coding agent with the paid model gateway OpenCode Zen.

Subscriptions and enterprise licenses: model providers bundle harness and models into per-seat or team plans. Claude Code is part of several Claude subscriptions, while OpenAI Codex is monetized through ChatGPT tiers, credits and API usage. Enterprise deals add identity management, audit logs and compliance features. How contested the segment has become shows in moves like OpenAI’s acquisition of the German startup Ona to strengthen Codex.

Managed runtime: running an agent outside your own devices requires isolated compute, persistent state, secrets management and logging. Microsoft’s Foundry Hosted Agents market that layer as a consumption-based runtime that scales to zero when idle.

Inference gateways: a model-agnostic harness can become a distribution channel for compute. OpenCode Zen and Nous Portal bundle models from various providers and bill usage through their own credit systems.

Security, governance and memory: permissions, policy checks, observability and long-term memory can be sold as standalone products. ProofAgent combines an open evaluation harness with a commercial governance platform, and Mem0 offers memory as a tiered cloud and enterprise service.

Marketplaces and vertical harnesses: paid skills and specialized agent environments for law, finance or medicine are plausible. Distribution channels such as the plugin marketplaces for Claude Code already exist, though a broadly established revenue-share model has yet to be documented.

Known Harnesses and Related Systems

The market ranges from minimalist open-source projects to comprehensive enterprise platforms. Some products serve harness functions without using the label themselves.

Some systems mentioned in the same breath belong in a different category. Nvidia’s NeMo Agent Toolkit describes itself as an open library that complements existing frameworks. Lefos, according to Earendil’s announcement, is an end-user application for email and Telegram built on top of the open Pi harness.

The Risks Sit in the Harness, Too

The more an agent is allowed to do, the more the boundaries matter. Compromised web content can steer an agent toward unwanted actions, and a badly scoped file permission can let it read sensitive information or alter data. Sandboxing, tool approvals and human confirmations are therefore core components of a harness. In the SHarD study, skill scanning, operating-system sandboxing and tool permissions improved the security scores of Claude Code and Codex substantially. Individual safeguards, however, collided with one another or were technically feasible in only one of the two systems.

Long workflows also remain error-prone, which weighs heavily where agents are meant to take over regular tasks in the workplace. For the DELEGATE-52 study, Microsoft Research examined 19 language models across 52 professional domains. Even leading models corrupted an average of roughly 25 percent of document content by the end of long sequences. In follow-up notes, the researchers stressed that they used a simplified harness and that the results transfer to production systems only in part. The harness thus cuts both ways: it accelerates an agent’s work while also being the place where errors get amplified, contained or caught.

A New Competitive Layer in AI

AI harnesses could become one of the most valuable layers in the emerging AI market, because they turn models into a workforce that can be used continuously. Their value comes from combining tools, memory, safety rules, model access and reusable workflows. Whoever controls that layer increasingly determines how an agent works and how easily it can move to another provider.

The base model keeps its weight in all of this. Research points to strong harness effects as well as to cases where the model has the larger influence. Successful systems will most likely come from a well-tuned combination of model, harness, tools and controls. For companies, that shifts the decisive question: what will count is which overall system works reliably over hours or days, keeps its costs in check, protects data, and can switch model providers when needed.