What Is an Agent Harness?

An agent harness is the software shell that turns a large language model into an autonomous actor. Without a harness, an LLM is a conversation partner: it answers questions, generates text, and follows instructions within a single session. With a harness, the same LLM can browse the web, execute code, manage files, interact with APIs, make purchases, and coordinate multi-step workflows across days or weeks without human intervention.

The term "harness" is precise. A harness wraps around a model the way a climbing harness wraps around a climber. It does not change what the model is, but it changes what the model can do. The harness provides the scaffolding for tool use, memory persistence, goal tracking, error recovery, and the ability to act on the world rather than just talk about it.

Agent harnesses are the actual autonomous actors in the agentic commerce ecosystem. Payment protocols define how money moves. Wallets hold the funds. Identity systems verify trust. But the agent harness is the entity that decides to search for a flight, compare prices across airlines, select the best option, and complete the booking. Without harnesses, the rest of the stack has nothing to serve.

This is why the Agent Harness category on the Agentic Commerce Market Map includes 12 companies spanning a wide spectrum, from coding-focused development tools to general-purpose autonomous assistants to crypto-native frameworks purpose-built for on-chain commerce. Each represents a different bet on what kind of autonomy the market needs most.

How Agent Harnesses Differ from Chatbots

The distinction between an agent harness and a chatbot is not one of degree; it is one of kind. A chatbot operates within a request-response cycle: you ask a question, it generates an answer, the conversation ends or continues. A chatbot cannot take actions in the real world. It cannot open a browser, fill out a form, click a button, or make a payment. It exists entirely within the text window.

An agent harness breaks out of this cycle. It has the ability to use tools, meaning software interfaces that let the LLM interact with external systems. A browser tool lets the agent navigate websites. A code execution tool lets it write and run programs. A payment tool lets it initiate transactions. A file system tool lets it read, write, and manage documents. Each tool extends the agent's reach from conversation into action.

Three capabilities separate agents from chatbots:

1. Tool use: the ability to interact with external systems (browsers, APIs, wallets, file systems)
2. Planning: decomposing complex tasks into sub-steps, choosing tools for each step, and adjusting when things go wrong
3. Persistence: maintaining state and memory across sessions, enabling multi-day workflows

The third capability is especially powerful. A chatbot's context disappears when the session ends. An agent harness maintains state across sessions, remembering what it has done, what it has learned, what tasks are pending, and what goals it is working toward. An agent can start researching flights on Monday, monitor prices through the week, and book when the price drops on Thursday.

Autonomy is the defining feature. A chatbot waits for human input at every step. An agent acts on its own, within the boundaries set by its operator, checking in with humans only when it encounters decisions outside its authorization scope.

The Autonomy Spectrum: From Coding Agents to Fully Autonomous Assistants

Not all agent harnesses offer the same degree of autonomy, and the spectrum reveals different philosophies about how much trust to place in AI systems.

At one end are coding agents, harnesses optimized for software development tasks. Claude Code by Anthropic, Cursor, and Devin by Cognition are prominent examples in the broader market. These agents can read codebases, write code, run tests, debug errors, and submit pull requests. They operate with significant autonomy within the development environment but stay tightly scoped to coding tasks. Their autonomy is deep but narrow.

In the middle are general-purpose autonomous assistants. Manus, OpenAI Operator, Project Mariner by Google DeepMind, and Perplexity Computer can browse the web, fill out forms, navigate complex multi-step workflows, and interact with virtually any web-based service. Their autonomy is broader (they can do almost anything a human can do in a browser) but they typically require human confirmation for high-stakes actions like payments or account modifications.

At the other end are crypto-native agent frameworks like ElizaOS and OpenClaw, purpose-built for on-chain autonomous commerce. These harnesses come with built-in wallet integration, x402 payment support, token management, and the ability to interact with DeFi protocols and smart contracts. Their autonomy extends to financial transactions: they can discover services, negotiate prices, and settle payments without human approval, within the spending policies defined by their operators.

The spectrum also reflects different safety philosophies:

• Coding agents limit risk by constraining the action space. The worst an agent can do is write bad code, which tests and code review can catch.
• General-purpose assistants add human-in-the-loop checkpoints for risky actions like payments or account changes.
• Crypto-native frameworks rely on programmable spending limits and on-chain guardrails. The wallet itself enforces the boundaries, not human oversight.

Coding Agents: Deep Autonomy in a Narrow Domain

Coding agents were the first agent harnesses to gain widespread adoption, and they illustrate the power of constrained autonomy. By limiting the action space to software development (reading files, writing code, running commands, interacting with version control) these harnesses can operate with remarkable independence while keeping risk manageable.

Claude Cowork by Anthropic represents the enterprise evolution of coding agents. Built on Claude's capabilities, Cowork extends the coding agent model into collaborative workplace tasks: not just writing code but managing projects, coordinating with team members, and handling the administrative overhead that slows down engineering teams. It bridges the gap between a pure coding agent and a general-purpose work assistant.

NemoClaw by NVIDIA takes the coding agent concept in a specialized direction, leveraging NVIDIA's deep expertise in GPU-accelerated computing and AI infrastructure. It focuses on the intersection of AI development and high-performance computing, enabling agents that can not only write code but optimize it for specific hardware configurations.

NanoClaw by Qwibit represents the lightweight end of the coding agent spectrum: a compact, efficient harness designed to run on consumer hardware rather than cloud infrastructure. This accessibility matters for developers who want local, private agent assistance without sending their code to external servers.

The relevance to agentic commerce is indirect but significant. Coding agents are building the infrastructure that every other agent type depends on. When a coding agent autonomously deploys a new x402 payment endpoint, fixes a bug in a wallet integration, or optimizes the performance of a trading bot, it is contributing to the agent commerce ecosystem at the infrastructure layer rather than the transaction layer.

General-Purpose Assistants: Broad Autonomy with Human Oversight

General-purpose autonomous assistants represent the most visible category of agent harnesses, the ones that non-technical users interact with directly. These are the agents that can do what you do in a browser: search for products, compare options, fill out forms, navigate checkout flows, and complete purchases.

Manus made headlines as one of the first truly general-purpose autonomous agents. It can browse the web, interact with any website, manage multi-step workflows, and handle tasks that span multiple services. Ask Manus to plan a trip and it will research destinations, compare flight prices, find hotels, check restaurant reviews, and present you with a complete itinerary, having navigated dozens of websites autonomously.

OpenAI Operator takes a similar approach but with the backing of OpenAI's GPT models and a focus on reliability and safety. Operator includes built-in guardrails for financial transactions: it can browse and research autonomously but pauses for human confirmation before making purchases or entering payment information. This human-in-the-loop model addresses the fundamental trust problem, that users want autonomous research but manual approval for spending.

Project Mariner by Google DeepMind integrates deeply with Google's ecosystem (search, maps, flights, shopping), giving it native access to services that other agents must access through web browsing. This integration advantage means Mariner can complete tasks faster and more reliably than agents that rely on parsing HTML and simulating clicks.

Perplexity Computer extends Perplexity's search capabilities into autonomous action. Where Perplexity the search engine finds information, Perplexity Computer acts on it: booking reservations, scheduling appointments, and completing purchases based on the information it has gathered.

Simular AI focuses on building agents that understand and interact with desktop applications, not just web browsers. This extends the agent's reach to software that does not have web interfaces, including spreadsheets, design tools, development environments, and enterprise applications.

MultiOn takes a browser-automation-first approach, building infrastructure that lets any AI model control a web browser with human-like precision. Rather than building a complete agent, MultiOn provides the browser interaction layer that other agent harnesses can use, making it a foundational tool in the ecosystem.

For agentic commerce, general-purpose assistants are the demand side of the equation. They are the agents that discover products, compare prices, and initiate purchases, driving transaction volume through the payment processors, wallets, and protocols that make up the rest of the stack.

Crypto-Native Frameworks: Agents Built for On-Chain Commerce

Crypto-native agent frameworks represent the most radical vision for agent autonomy. Unlike general-purpose assistants that pause for human approval before spending money, crypto-native agents are designed to transact autonomously, discovering services, negotiating terms, and settling payments on-chain without human intervention.

OpenClaw is the most prominent open-source agent harness built specifically for autonomous commerce. It runs locally on the user's device, maintains its own wallet, and can browse the web, interact with APIs, and make payments using x402 and stablecoins. OpenClaw's design philosophy prioritizes user sovereignty: the agent runs on your hardware, using your wallet, under your control. However, security analysts have noted that OpenClaw is "insecure by default," meaning users must add their own safety layers (like virtual debit cards) to prevent unauthorized spending.

IronClaw is a hardened variant of the OpenClaw framework, focused on security and enterprise deployment. Where OpenClaw prioritizes openness and flexibility, IronClaw adds mandatory spending controls, encrypted credential storage, and audit logging, providing the guardrails that enterprise deployments require.

ElizaOS is the agent operating system built by the ai16z community, designed from the ground up for crypto-native autonomous operation. ElizaOS provides a full-stack environment for building agents that can interact with DeFi protocols, manage token portfolios, execute trades, participate in governance, and engage in agent-to-agent commerce through Virtuals Protocol's ACP. With over 18,000 agents running on Virtuals Protocol and a combined aGDP exceeding $470 million, ElizaOS-powered agents represent the largest autonomous agent economy operating today.

The significance of crypto-native frameworks for agentic commerce cannot be overstated. They are proof that autonomous agent commerce works in practice, with real money changing hands, real services being purchased, and real value being created. The challenges they face (security, trust, regulatory compliance) are the challenges the entire ecosystem will need to solve as agent autonomy increases.

What Agents Actually Do: Browse, Transact, Code, Coordinate

Agent harnesses are defined by their actions, and those actions fall into four broad categories that map directly to agentic commerce use cases.

1. Browsing: Agents navigate websites, read content, fill out forms, click buttons, and interact with web applications. For commerce, this means searching for products, comparing prices, reading reviews, and navigating checkout flows. General-purpose assistants like Manus, OpenAI Operator, and MultiOn excel here.
2. Transacting: Agents make payments, manage wallets, interact with smart contracts, and settle financial obligations. Crypto-native frameworks like ElizaOS and OpenClaw are purpose-built for this, with native wallet integration and x402 support.
3. Coding: Agents write, test, and deploy software, including the software that powers other agents. Coding agents like Claude Cowork and NemoClaw are continuously improving the tools and platforms the rest of the agent economy depends on.
4. Coordinating: Agents interact with each other through protocols like A2A and ACP, discovering other agents, delegating tasks, negotiating terms, and collaborating on complex workflows.

These four actions are not independent; they combine in complex workflows. A shopping agent browses to find products, coordinates with a price comparison agent, transacts to make the purchase, and a coding agent maintains the infrastructure that makes it all possible. The agent harness is the orchestration layer that ties these actions together.

The Competitive Landscape

The agent harness market is one of the most intensely competitive spaces in AI, with well-funded companies and major tech platforms all racing to build the dominant agent framework.

The competition breaks along three fault lines:

• Open versus closed: OpenClaw and ElizaOS are open-source, letting anyone inspect, modify, and extend the code. OpenAI Operator and Project Mariner are closed-source, proprietary products. Manus occupies a middle ground with selective openness.
• Local versus cloud: OpenClaw runs entirely on the user's device, with no data leaving the machine unless the agent explicitly sends it. Most other harnesses run in the cloud, with the agent's state, memory, and actions managed by the provider.
• General versus specialized: Manus and OpenAI Operator aim to handle any task a human can do in a browser. ElizaOS specializes in crypto-native operations. Claude Cowork specializes in software development. NanoClaw targets resource-constrained environments.

The market has not yet decided whether the winning approach is one agent that does everything or a network of specialized agents that collaborate. Each of the 12 companies in the Agent Harness category represents a different bet on which combination (open/closed, local/cloud, general/specialized) will win. The answer is likely that multiple approaches coexist, serving different user segments and use cases, with coordination protocols like A2A enabling interoperability between them.

How Agent Harnesses Connect to the Rest of the Stack

An agent harness does not operate in isolation. It sits at the center of the agentic commerce stack, connecting to every other layer.

At the model layer, the harness wraps a foundation model (Claude, GPT, Gemini, Llama, DeepSeek, or others). The model provides the intelligence (reasoning, language understanding, planning), and the harness provides the capabilities (tool use, persistence, action execution). The choice of model affects what the agent can do, but the choice of harness affects what the agent is allowed to do.

At the protocol layer, the harness implements the standards that enable interoperability. MCP support lets the agent discover and use tools. A2A support lets it coordinate with other agents. x402 support lets it make payments. ACP support lets it navigate shopping flows. The more protocols a harness supports, the more of the ecosystem it can access.

At the wallet layer, the harness connects to the agent's financial infrastructure. For crypto-native harnesses, this means direct wallet integration, where the agent can sign transactions, check balances, and manage funds. For general-purpose harnesses, this means integration with payment APIs, virtual debit cards, or human-in-the-loop approval flows.

At the identity layer, the harness carries the agent's credentials: ERC-8004 registration, KYA verification, and reputation scores. These credentials determine which services the agent can access and how much trust it is granted.

At the discovery layer, the harness uses directories, marketplaces, and indexes to find services and other agents. Without discovery, an agent can only interact with services it has been explicitly configured to use. With discovery, it can autonomously find new services, evaluate them, and expand its capabilities.

This multi-layer integration is why building an agent harness is so difficult. It is not enough to wrap an LLM in a tool-use framework. A production-grade harness must handle payments, identity, discovery, error recovery, security, compliance, and coordination, all while maintaining the conversational intelligence that makes the agent useful in the first place.

Key Companies Building Agent Harnesses

The 12 companies in the Agent Harness category each bring a distinct approach to the market.

OpenClaw is the leading open-source agent harness for autonomous commerce, running locally on user devices with built-in wallet integration and x402 payment support. Its open-source nature makes it the most inspectable and customizable option, though users must add their own security layers. IronClaw extends the OpenClaw foundation with enterprise-grade security: mandatory spending controls, encrypted credentials, and comprehensive audit logging.

ElizaOS by ai16z is the agent operating system for the crypto-native ecosystem, powering the largest autonomous agent economy through Virtuals Protocol. With native support for DeFi interactions, token management, and agent-to-agent commerce via ACP, ElizaOS represents the most fully realized vision of on-chain agent autonomy.

Manus is the breakout general-purpose autonomous assistant, capable of browsing the web, managing multi-step workflows, and interacting with virtually any web-based service. Its combination of broad capabilities and reliable execution has made it one of the most widely used agent harnesses.

OpenAI Operator leverages GPT's capabilities with a focus on safe, reliable autonomous browsing and task completion. Its human-in-the-loop model for financial transactions positions it as the trusted choice for users who want autonomy with guardrails.

Project Mariner by Google DeepMind integrates with Google's service ecosystem for faster, more reliable task completion. Perplexity Computer extends search intelligence into autonomous action. Simular AI focuses on desktop application interaction beyond web browsers. MultiOn provides browser automation infrastructure that other harnesses can build on.

Claude Cowork by Anthropic bridges coding agents and workplace assistants, bringing autonomous capabilities to enterprise collaboration. NemoClaw by NVIDIA specializes in AI development and high-performance computing workflows. NanoClaw by Qwibit offers a lightweight harness that runs on consumer hardware, making agent capabilities accessible without cloud infrastructure.

The Future of Agent Harnesses

The agent harness category will evolve in three important directions over the next several years.

First, convergence between categories. Today, coding agents, general-purpose assistants, and crypto-native frameworks are distinct products built by different teams. The future points toward unified harnesses that can code, browse, transact, and coordinate, switching modes based on the task at hand. An agent that can write code to deploy a new MCP server, browse the web to find customers, negotiate pricing through A2A, and settle payments via x402 is more valuable than four separate agents that each handle one piece.

Second, the safety and security infrastructure will mature. Today's harnesses range from "insecure by default" (OpenClaw) to "overly cautious" (requiring human approval for every action). The future requires graduated trust: agents that can make small purchases autonomously, escalate to human approval for large transactions, and operate within programmable policy boundaries that adapt as the agent builds a track record. Virtual debit cards, programmable wallets, and on-chain spending limits are the early building blocks of this graduated trust model.

Third, specialization through composition. Rather than every harness trying to do everything, the ecosystem will likely settle on a pattern where lightweight, specialized agents collaborate through A2A and ACP protocols. A shopping agent delegates payment to a payment agent, which delegates identity verification to a trust agent, which delegates wallet operations to a wallet agent. Each agent is simple and well-tested. The complexity emerges from composition, not from monolithic harnesses trying to handle every edge case.

The companies building agent harnesses today are building the robots of the digital economy. Not physical robots that move through factories, but software robots that move through the internet, browsing, buying, building, and coordinating with a level of autonomy that would have been science fiction five years ago. The harness is where the intelligence meets the action, and the 12 companies in this category are defining what that intersection looks like.