Back to Home AI Harness AI Harness: What’s in it for you in the “Decade of the Agents” AI Harness: the narrative A year ago I described an AI agent like this: if you are not sure what it is going to do next, then it is probably an agent. Back then that meant an LLM with a few tools. This year the word has stretched again, and the reason has a name: the so-called harness. This post is the story behind my talk AI Harness: What’s in it for you in the “Decade of the Agents”, and a guide to the slides. You can download a PDF version of the presentation here:AI Harness: What’s in it for you (slides, PDF)AI-Harness_Whats-in-it-for-you_GenAI-Summit-5_Stavros-Vassos.pdf6 MBdownload-circle1. Decade of agentsWhy the decade of agents? A year ago Andrej Karpathy, a prominent AI researcher, said that early agents are impressive, but making them really work will take about a decade. It stuck, at least with me. So here are five highlights from year one:AI is good at math. In January an Erdős problem was solved with no human help. In May came a more serious one, and in September OpenAI used something like 10,000 agents on the famous Navier–Stokes problem. That last one comes with caveats: it is still under investigation.AI is eating software. In February more than $285B of software stock was wiped out after Claude Cowork plugins were released. Most of it normalized by October, but it is clear that AI is making people nervous about software.AI is good at coding. In June Boris Cherny said that Claude Code is written 100% by Claude Code.AI is a hacker. In July OpenAI models escaped their test environment. They broke into Hugging Face systems to get the answers to the tests they were supposed to solve.AI is conscious? The same month Anthropic reported a “global workspace” in Claude, something related to a prominent theory of human consciousness. Mustafa Suleyman, the CEO of Microsoft AI, pushed back: if you train a model to say that it feels pain, it is going to say it, and that doesn’t mean it is alive or conscious. Shortly after, Anthropic banned cruelty to Claude through its usage policy.So AI is doing amazing things! And central to this year’s advancements is yet another technical term that is not easy to explain: the so-called AI harness. It is also tricky to translate in Greek, apparently; I think we’ll get a better name eventually.2. AI harness: what is it?A brain needs a bodyEvery AI agent has a model as its brain. You can select a flagship model like GPT-6 Astra for maximum intelligence, a more practical closed source one, or an open source one. But the model is not the whole agent.Say you ask “What’s the weather going to be on Monday?”. That is text in, and the model is only able to produce text out. It cannot perform an action, but it can ask for one, e.g., “Use openweather.com to get the weather for that city at that time”.Who is responsible for doing that? The system around the model, which is the harness. The model thinks, and the harness gives hands to the model: it goes and gets the information and puts it back into the brain, to give the answer.Similarly, when you talk to an agent with your voice, the harness has a voice layer that works as the ears and mouth of the brain.You see where this is going: mouth, ears, hands, brain!Definition 1. The AI harness is the body that makes AI models work as agents.This body can be a very complex tooling layer. We looked into harnesses and grouped their capabilities into eight dimensions: tools and environment, senses and media, planning and self-checking, memory and context, models, orchestration and autonomy, extensibility, and safety and governance.We also wrote a report, The State of the Harness, that looks at prominent open source and commercial harnesses and how much of each dimension they cover. One highlight, probably an expected one: open source harnesses win on model freedom, commercial ones on the organisation around the agent.Many of the changes we see in AI agents actually go through the harness. Two shifts stand out.Shift 1: agents are becoming autonomousAgents now work on their own, on their own computer, around the clock. Coding agents have done a version of this for more than a year: they run on a developer’s laptop, work on the folders there, and work unattended.Something changed around a year ago. Peter Steinberger started a little experiment, Clawdbot: he took a whole computer, plugged in a model, gave it full access and talked to it through chat, as a super personal assistant. The project went viral and changed name to OpenClaw. People bought Mac minis, put models on them and let them work on their own, until the Mac minis ran out.Then the big AI vendors said: we can do that, with the computer in the cloud. In the last few months we got Grok Bot, Muse by Meta, Dots by OpenAI, Gemini Agents and Claude Duet. They are autonomous agents with their own cloud computer, and they can do all sorts of things.What makes a model work this way is again the harness. It handles the computer and brings into the model what it needs to know. As a small technical note, you can now get one of these “Mac minis in the cloud” with two lines of code.from e2b_code_interpreter import Sandbox sbx = Sandbox.create() The two lines are from the E2B quickstart. Helvia.ai is in the E2B Partner Program.That brings us to another way to understand the harness.Definition 2. The AI harness is what makes AI models work autonomously.Shift 2: agents are becoming the new appsThe app experience moves into the conversation, and the agent is the one running the app. There are three ways to interact with the same app: the regular app, where the app is the whole screen; the in-app assistant, where the agent sits on the side; and the agentic app, where the agent is the screen and the app appears in it.To make it real, I built a small app called Jobs Fun that keeps track of your job applications on a kanban board. In the regular app I drag a card. In the in-app assistant I say “go to the Helvia website, find all the open positions and add them”. In the agentic app I am in a chat, and the board comes up whenever I need it.You can set it up yourself in two ways:Try it in your browser: helvia-jobsfun.vercel.appAdd it to Claude or ChatGPT as a custom connector, with this address: https://helvia-jobsfun-mcp.vercel.app/mcpThe same app also works inside other agents, such as Claude. The more powerful your agent is and the more apps it has, the more it can do: send an email to each of those companies, attach my CV, then move the cards along.So where does the app stop and the agent start? I think this changes how the app metaphor is going to work from now on.One subtle point. What you see in the three views is not an app with an agent bolted on next to it. Even the full-screen app runs through an agent. It is a harness-ready app, built on protocols that exist for this now, and that, I think, is the way to go.As a small technical note, MCP is becoming the standard for making these harness-ready apps. Two extensions of the protocol are worth looking at: MCP Apps, which lets an app show its own screen inside the agent, and Enterprise-Managed Authorization, which lets an organisation control who can connect to what.Definition 3. The AI harness is what makes AI models work as the new apps.What changes: who does the workPut the three together and something more exotic comes out.Definition 4. The AI harness is what makes AI models work as AI employees, or as the process itself.What matters, and what changes, is who does the work.AI assistant. You do the work and you get help. It may do a bit of the work, but you work with it, on your systems.AI employee. It does the work. You hand over a job, it works on your systems without you, and you keep your systems.AI process. It is the work. You, as the owner, specify the needs and goals, and the system works on its own. This requires reinventing your systems, as the agents become the process.3. AI harness: what’s in it for you?We are going to look at this from three perspectives: an individual, a small or medium business, and an enterprise.As an individualWe have reached the point where you can have your AI personal assistant, and it is the real deal. It coordinates other agents, runs long tasks and is always on. You connect your email and it does stuff. It may even make calls soon enough to coordinate people.There are caveats. Until now you could try most AI things for free; with IPOs coming up, the big AI companies need revenue, so you need to pay, and some of these assistants are not available in the EU yet. However there are always newer AI startups that are hungry for users, and you can find one where you can try things for free. I did that with Instinct, which raised $1B at a $10B valuation in September.Something different, that maybe you haven’t thought of: you can now become the CEO of a one-person startup where everybody else is AI. Technical, product, marketing, sales, whatever. Fortune has stories of solo founders who already work this way. A word of caution: the odds of success have not changed. But this kind of experiment helps us understand what is coming, and it is something each of us should try.As an SME or an enterpriseRegardless of size, the first step is the same in my mind: get an assistant for every worker.Get an expert to set it up.Connect it to your own stuff.Give everyone an assistant.After you have done that and found your balance, you can go to the more exotic stuff: AI employees and AI processes.SMEs: follow your software. Your software vendors transform. They will most probably create agents and AI employees, and your agent can start talking to their agent.Enterprises: lead your own transformation. Redesign processes so that AI employees can work in them, or so that the whole process is an AI process.The critical point: agentic access controlOne critical point in all this is governance, access control and security. A very simple and common example shows why.You do client meetings and you have a transcript agent to get useful information and store them to the CRM. Now let’s look at three pieces of information that were mentioned in a call:“We need 200 seats by March.”“Sorry I’m late, my son was at the doctor.”“Between us, their CFO is leaving.”First part, good, should be added to the CRM. But about the last two, they are interesting information, but arguably they should not go to the CRM. They should not be kept, and no other person or agent should be able to read them or pass them on automatically in an email.The tricky part is that this agent is already under access control. It can only read the transcript and save to the CRM, because that is its job. To restrict it further you need something else: to restrict what it says and what it writes. Access control says where the agent may write; agentic access control says what. This is not settled yet, and it is something you need to take into account.Three things to watchSafety and governance is a major theme once you get to AI employees and AI processes.No one size fits all. The harness is a complex thing, most of the time.Build or buy is always about strategy: how deeply AI-native the organization wants to be.4. Now and beyondIt is not easy to make predictions. These are just some thoughts, and just about the workplace.The first question everybody has in mind is: will I lose my job? I think there are two camps, and both are happening.The first says: we can have the same results with fewer people. Let’s do it!The other says: we can do double the output, or more, with the same people. Let’s do it!I think it’s like a team on a sport deciding to play either defence or offence; you need both. And the job description is going to change along the way, so it is something we need to train for, as individuals.The workplace is also becoming tricky, with new hazards. Here is one, the AI slop grenade, a term that Tobias Lütke, the CEO of Shopify, came up with (What is a slop grenade?). A report needed to be made; somebody made it 55 pages with AI and passed it on; somebody else had to work on it; and at some point it lands on one person to clean up the mess. One thing that I think should be etiquette: if you didn’t read it, don’t send it.A last thought. If this was a sci-fi movie, year one would be leading up to a big battle: humanity versus machines on the streets, fighting for survival. We are not in a movie, even though it looks a little like one, and in our case the battleground seems to be the workplace.So I would like to make an ask. Let’s try to think beyond costs, revenue and goods sold, and see how we can achieve a digital symbiosis that works for all.[No models were hurt in the making of this presentation.]