The AI Harness Is Where Enterprise Software Competes Now

In one quarter, three companies started using the same word. Adobe's documentation for CX Enterprise Coworker lists an "enterprise harness" as one of the product's four building blocks. WRITER announced in August that it had rebuilt the "WRITER Agent harness" to run "regardless of the underlying model." Then on September 11, Salesforce named its entire agent architecture the Trusted Enterprise AI Harness. I believe the vocabulary is telling you where these companies think durable value sits and it's not at the model layer.

A disclosure before going further. I currently sell in Adobe's Digital Experience business, so I see one of these harnesses up close every week. Everything here is my own reading of public documents, analyst notes, and announcements, and none of it speaks for Adobe.

A harness is everything in the agent that is not the model

The cleanest definition I have found comes from Lang Chain's engineering blog: "If you're not the model, you're the harness." Adobe's own documentation is more specific about what that means in practice, describing the enterprise harness as "the engine that turns your goal into completed work. It runs a continuous, self-correcting loop (reason, select a skill, execute, validate, repeat), recovering from errors and working in parallel until the job is done, not just attempted."

The model does the reasoning while the harness does everything else: it picks the playbook, calls the system, checks whether the result matches the goal, and decides whether to go again, stop, or hand the decision to a human. A chat window is a model with no harness, which is why you have to check its work yourself. A coworker is a model inside a harness that checks the work before you ever see it.

Most of the confusion I hear about these products, including from people who sell them, comes from collapsing the layers.

Validate is the step that separates a coworker from a chatbot

Of the four verbs in the loop, three produce something you can watch, while validating produces nothing visible when it succeeds, which is why it gets the least attention in demos and why it is the step that decides whether an enterprise can let the loop run at all.

Biology solved this problem long before software did. Your body holds its temperature within about a degree through a feedback loop that senses, compares against a set point, and corrects thousands of times a day without a meeting, and the cells doing the correcting do not choose the set point. A harness is that loop applied to a business process, governance is the set point, and the customer rather than the vendor gets to choose it.

Nick Patience of The Futurum Group reported in July that Adobe executives describe a four-level autonomy model and place most of the customers they talk to at level one or two, with the near-term goal of moving them to three, and that Adobe cited on stage that only around a quarter of enterprise AI proofs of concept have ever reached production. McKinsey's 2026 survey lands in the same place from the other direction, with about two in ten organizations reporting they have reached the scaling phase with AI agents.

So the harness can run at level four, and most companies have it set to two, which means the ceiling is being set by the organization. That is why a better harness makes a badly governed company worse, faster: the loop amplifies whatever set point it is given, so sloppy permissions at level one produce a bad report while the same permissions at level three produce a bad campaign that already went out. (Also read: Enterprise AI Is an Organizational Design Problem in Disguise)

The Self-Correcting Enterprise is what you get when the loop is set right

Here is what the current cycle looks like at most companies I work with. A campaign runs for a quarter, someone pulls the dashboard before the business review and notices the offer underperformed for one segment, and the fix ships in the next planning cycle. The correction cycle is ninety days because the humans who can authorize the change meet every ninety days.

I am going to call the alternative The Self-Correcting Enterprise: a business whose correction cycle runs at the speed of its harness loops rather than the speed of its meeting calendar. The loop senses the miss on day two, drafts the fix, checks it against the governance rules the company set, and either ships it or hands a one-line decision to a human before lunch. The company's judgment did not change; it was encoded once, in the set point, and applied continuously instead of quarterly.

Jet2 is the closest public example so far. The airline signed a multi-year deal with Adobe on September 22 to run CX Enterprise Coworker across 10 million myJet2 customers, with offers shaped by travel preferences and booking history and daily itinerary suggestions once a customer is on the ground, so where a brochure corrects itself once a season, this loop corrects itself every morning. Adobe is embedding forward-deployed engineers in a joint lab to make it work, which says something about how much of this is still craft. (Also read: Inside the Engine Room of Enterprise AI: Why Forward-Deployed Engineers Matter)

Three kinds of harness are competing, and they wrap different things

Once the loop is a shared pattern, the vendor fight reorganizes around what each harness wraps and where its set point comes from.

The most revealing line in this whole quarter came from Salesforce's own president of platform, Rohan Kumar, in the harness announcement: "Models will continue to change, and intelligence will increasingly be available everywhere. What will differentiate an enterprise is the trusted, proprietary context it brings to that intelligence." That is a CRM vendor conceding that the loop itself is a commodity and betting the company on what the loop checks against. Adobe is making the identical bet with a different noun, where the context is audiences and campaigns instead of accounts and pipeline. (Also read: What Survives When AI Becomes a Commodity)

The point solutions are betting on something else, and their bet is underrated. A support harness has a narrow, measurable target, so it can validate against resolution rate and run at a higher autonomy level sooner, which is how Decagon can claim average deflection above 80 percent and Sierra can count more than 40 percent of the Fortune 50 as customers. Narrow loops validate faster than broad ones, so the startups may reach real level-three autonomy in production while the suites are still moving customers off level two.

The suites' counter is sprawl. A VentureBeat Intelligence survey this July found 85 percent of enterprises already run two or more agent orchestration platforms, averaging 3.1, and nobody wants three set points. The harness that wins is the one whose governance the others inherit, and inheriting permissions from the system that already holds the customer data is an advantage.

Where this goes over the next eighteen months

The question I keep circling back to is what a company looks like once its correction cycle is an hour instead of a quarter. The org chart most enterprises run today exists largely to move decisions up to the people allowed to make them, on a schedule those people can attend, and a harness with the set point encoded once removes most of that traffic. Within a couple of years the loop will be the same everywhere. The set point is the company.

Key Takeaways

• An AI harness is the loop around the model that plans, acts, checks, and stops; the model reasons and the harness makes the reasoning safe to act on.
• Autonomy is a customer setting, and most enterprises currently run Adobe's four-level model at one or two, so the ceiling on agentic value is organizational before it is technical.
• Validate is the step that makes a harness trustworthy, and a better harness makes a badly governed company worse faster because the loop amplifies whatever set point it is given.
• The Self-Correcting Enterprise is a business whose correction cycle runs at loop speed instead of calendar speed, with judgment encoded once in governance and applied continuously.
• Adobe, Salesforce, and the agent-first startups run the same loop; they compete on what the loop wraps and where its governance comes from, and narrow loops validate faster than broad ones.
• Enterprises averaging three orchestration platforms will consolidate around the harness whose governance the others inherit.

FAQ

What is an AI harness? An AI harness is every part of an agentic system that is not the model: the loop that reasons, selects a tool or skill, executes, validates the result, and repeats until the goal is met or a human needs to decide. The model supplies the intelligence; the harness supplies state, tool access, error recovery, and the approval gates that make autonomy safe.

Is Adobe CX Enterprise Coworker an AI harness? Coworker is the product, and Adobe's documentation lists an "enterprise harness" as one of its four building blocks alongside skills, MCP connectors, and governance. The harness is the engine inside Coworker that turns a goal into completed work through a self-correcting loop.

What is the Self-Correcting Enterprise? A business whose correction cycle runs at the speed of its agent harness loops rather than the speed of its planning calendar. Judgment is encoded once in governance rules and autonomy settings, then applied continuously, so a missed target is corrected in hours instead of at the next quarterly review.

About the Author

Martin Goetzinger has spent his career in enterprise software sales, helping large organizations such as Apple, Microsoft, and Verizon connect data, insight, and action. His work focuses on transforming how businesses measure success and create customer value through technology.

Outside the enterprise world, he writes about the five forces he believes are reshaping everything: AI, blockchain, energy, personalized health, and robotics. Not from a purely technical lens, but from a human one as to how these technologies will redefine work, wealth, and well-being.

He is based in the U.S. and publishes at www.MartinGoetzinger.com.

Disclaimer

The views expressed in this article are the personal opinions of the author and are provided for informational and educational purposes only. Nothing in this article constitutes investment advice, financial advice, legal advice, or any other form of professional advice. Do not make investment or financial decisions based on the content of this article. Always consult a qualified professional before making decisions that affect your finances, business, or livelihood.