Auto harness: Your product advantage

A question I keep getting: what's the advantage of a software product when AI models are getting incredibly good at programming, and the frontier is moving to the application layer? You can literally use Claude Code or Codex for a lot of knowledge-worker tasks beyond coding — accounting tasks, finance tasks. So what's left to defend? Why are people still building "AI for X" (law, accounting, etc.)?

It turns out building the harness around the model is the actual advantage. Average people won't spend a ton of time setting up the integrations, skills, infrastructure, and human-in-the-loop process around the application. You can leverage Claude Code or Codex for the agentic loop — that part is easy now. The hard part is everything around it.

Say you're building an AI agent for software engineers. The harness is the sandbox environment, the provisioned database, analytics, and the resources it needs like queues. Obviously experienced engineers can use Claude Code to set all this up themselves — give it AWS credentials and ask it to create the cloud resources, and it does a decent job. Wire in logs and analytics so the agent has context to fix its own issues, then have it document what it did. It's a loop.

And the best harnesses close that loop on themselves. When the agent makes a mistake, it shouldn't just fix it — it should record it and learn from it. Fix a recurring bug, then write a new skill so it doesn't happen again. Resolve a common review comment, then update CLAUDE.md so the next run already knows. The harness documents its own errors and gets better every time it runs. That auto-learning is the part Claude Code out of the box won't do for you — and it's the part that compounds.

Another use case: an AI agent for tax preparation. Think about putting in all your financial data — invoices, bank statements — connecting all your financial accounts, running the tax calculation, and getting a return back with the data. But the better version shows, on the return itself, how every number was calculated and where it came from, so people can actually review it. It has to be auditable. And it has to file with the government and track the response.

The same auto-learning loop applies here too. When a return gets corrected on review, the system records the fix and updates its own rules — so the next return is more accurate, and the one after that. Every reviewed return makes the harness smarter, the way a senior accountant gets sharper with every filing season.

All of these things are about building the harness around the agentic loop.

The software-engineer and tax-accountant examples are for professionals. Now think about non-technical people — how are they going to use Claude Code to provision infrastructure and build the harness themselves? They aren't. And non-accountants — how do they review a tax return file? Dedicated tax AI software saves them far more time. That's a 10x–100x bigger market, because there are far more non-engineers than engineers.

Now someone will push back: if the harness is just integrations and infrastructure, a funded team copies it in a quarter. True. The scaffolding is table stakes — anyone can build it, and eventually everyone will.

But the scaffolding was never the advantage. The compounding is. The skills the harness writes for itself, the rules it updates after every fix, the corrections that make each run better than the last — that part isn't built, it's accumulated. You can copy my setup. You can't copy a year of my harness teaching itself from real usage. That's the gap that keeps widening while everyone else is still wiring up their sandbox.

When you build a software product for legal, accounting, marketing, etc., you've already put the knowledge inside it. That's the business domain. You design and test the prompts with tool use, and you provide the harness environment that normal people would take weeks or months to set up — and that keeps improving itself the more it runs.

Build your own auto-harness environment for your product. It compounds over time. That is your advantage.