AI Agents / POV

If Agent Skills Are Easy to Copy, Where Is the Moat?

Skills and agent workflows are useful, but are they a business moat? I think the stronger opportunity is a harness that understands how a specific domain works.

On this page
  1. What happens when the model gets better?
  2. My answer: a sector-specific harness
  3. Why would this be harder to copy?

I was talking to a 26-year-old Indian founder in Silicon Valley on Friday. He is building a system to find security vulnerabilities in a codebase. As I asked how it works, I found that much of the system, as he described it, is a combination of agentic workflows, mostly built around skills.

I have noticed this with many founders. Someone builds skills for marketing. Someone else builds them for insurance. The model receives instructions, follows a workflow, uses a few tools and produces an answer. Some of these products solve real problems. But I keep coming back to the same question: what is the business moat?

Knowing the domain and improving a workflow with clients on live projects both help. Still, if the product is mainly a set of instructions based on which a model reasons and acts, I find it hard to see the lasting advantage. Another team can write similar instructions. More importantly, the model may eventually need fewer of them.

What happens when the model gets better?

BMad is an interesting example. It has built a substantial set of skills and workflows for software development. Many of those workflows are powerful. But coding harnesses such as Codex and Claude Code, and the models inside them, keep improving. Some guidance that was essential earlier may become less useful after a model or harness upgrade. OpenAI has made a similar observation about elaborate skills and prompts.

That does not mean BMad has no value. My question is about any business built primarily on workflows: how do you know which parts will still be needed after the next upgrade? If the instructions can be copied and some can become obsolete, is the workflow itself enough of a moat?

Could fine-tuning a model for a specific use case be the answer? It may save tokens, improve a narrow task or allow self-hosting. But training a model does not remove the need to integrate it, keep it current and check whether the whole workflow improves. I have explored this separately in Where Could Specialised Models Fit in an Agent Workflow?. I do not think fine-tuning, by itself, settles the moat question.

My answer: a sector-specific harness

I think the stronger opportunity is to build a harness that understands how work happens in a particular sector. It should know which skills and workflows to use, but also which systems contain the relevant information, what can be trusted and when a person needs to make the decision.

Memory is a large part of this. What should the system remember from the current task? What belongs in a client’s long-term context? Which previous cases are useful, and which are dangerous to reuse because a rule or document has changed? When should the harness retrieve information, and how much context is enough for the model to act accurately?

This is where domain taxonomy and ontology can help. They give the harness a way to understand the terms and relationships that matter in that sector. An insurance policy, an endorsement, a claim and a decision are connected, but they are not interchangeable pieces of text. A marketing claim, the evidence supporting it and the person allowed to approve it also have different roles. That structure needs to help the system find the context needed for a decision. Otherwise, it is more work to maintain.

Why would this be harder to copy?

Take a hypothetical insurance claim. A policy was issued in January, its cover changed through an endorsement in March, and a claim arrived in April. A generic skill might read the first policy document it finds and produce a convincing recommendation. A domain harness would need to find the version that applied when the loss occurred, check the endorsement, identify what the evidence does not establish and send an uncertain case to the right reviewer.

What happens after the reviewer corrects it? Was the wrong document retrieved? Did the system miss an effective date? Did it treat an unresolved fact as certain? If those corrections become tests and changes to retrieval or escalation, the next similar claim can be handled better. Over many real cases, the company builds knowledge of failures and of the customer’s actual decision process. A competitor can copy the skill much faster than it can acquire that learning and earn the same place in the workflow.

Of course, merely calling this a harness does not make it a moat. Integrations can be copied too, and client data cannot simply be shared across customers. We would need to show that the system makes better decisions, keeps up with changes in the domain and works reliably for the customer.

Harvey is a useful legal example of the direction I mean. Its Vault product page describes document analysis connected to its Assistant and Workflows. I cannot tell from a product page what Harvey’s moat is. But it shows why a legal system needs to work with the documents and processes around an answer, rather than only follow a legal prompt. I can imagine a similar kind of harness for insurance or marketing.

LLMs are already powerful, but they are general-purpose. For a specific business problem, their usefulness depends heavily on what they know at the moment they act. The opportunity, in my view, is to build the system that knows what to bring to the model, when to bring it and how to check what happens next. That is where a sector-specific moat may develop.