Yuechen Zhao
Essay · October 21, 2026

Stop Selling Agents Like Software

Build around what the models can do, and every release makes your product better for free.

For the past year, we built and sold AI agents like software. "Agent X does Y for Z / mo, half of what you're paying today!"

For a lot of startups, that was the right call. It let you deliver a reliable service for a well-defined set of tasks, with a compelling ROI story for the customer.

But it's time for all of us to rethink this paradigm, because a new era for agents has arrived: as models get ever more capable and get ever cheaper per token, "constraining AI" will be one of the biggest mistakes founders can make. Build around what the models can't do, and every model release erodes your moat. Build around what they can do, and every release makes your product better.

The end of the software era of agents

Behind the scenes, every agent was a carefully maintained specialist with a human babysitting it. Leave it alone too long and it would drift off track, or break something important.

Writing skill files became the new software engineering, and plenty of vendors took advantage: find the most expensive workflows, build the skills and tools, deploy carefully, charge a heck of a lot for it, "update" the agent like software after running evals, then upsell the next workflow as a new SKU.

That era is now behind us.

What changed

Personal agents got usable.

A personal agent is the hardest deployment there is. No engineers, no evals, no reviewed skill files. That's exactly what made everybody give up on OpenClaw in January. You had to be an engineer to keep it alive.

Now ordinary people set them up, correct them in plain English and hand them more over time. The agent figures out where feedback belongs, remembers it and applies it next time. Nobody engineered that. The models got good enough to be managed instead of programmed, and that's why personal agents are going viral.

Personal agents showed us who can run them: everyone. If a non-technical person can train an agent at home, a non-technical team can train one at work.

Make no mistake: agents will need more training than ever, because they're going to take over most of what people actively do. What changes is who does it, and when: everyone who works with the agent, every day, instead of the vendor's engineers before launch. You don't program a good employee. You manage one.

Don't limit what your agents do

If anyone can keep handing an agent new work, selling it one workflow at a time stops making sense. Devin and Clay already sell plans that are really pools of metered work. Even Figma, the poster child of seat-based SaaS, is moving its new agent onto credits. For a long time I didn't get why that beats workflow pricing. They looked like two routes to the same place.

They're not.

SaaS was expensive to build and nearly free to run, so you charged per capability. Agents flip that: teaching one a new job costs almost nothing, but doing the work costs real money. Software gets sold by the feature. Workers get paid for work.

Selling by workflow also kills what makes agents valuable: customers training them. If they want the agent to do one more thing, they have to buy the next SKU. An agent nobody can train never takes on more work, so it never delivers more value. That's a death sentence.

Think about it: your best employees don't stay in one lane. As they learn more about your company, they take on more, and you give them more. Agents should work the same way. Let customers hand your agent more, and charge them more as it delivers more.

Trust is part of the product

"Don't limit your agent" doesn't mean "let it run wild." It means your engineering moves. Instead of building SKUs that spell out everything the agent can do, build the guardrails that spell out what it can't: permissions, authentication, audit trails, and hard walls around the systems it can't touch, the money it can't move and the catastrophes you can't risk. That's not overhead. It's part of what customers pay you for, because nobody hands real work to an agent they can't trust.

Then loosen them as the agent earns it, the same way a new hire works their way up to signing authority.

Do more, own more, be accountable for more

As agents earn trust, they earn responsibility, and pricing will follow the same way pay does. An associate bills by the hour, a senior gets a bonus, an executive gets a share of profits. Today, most agent companies charge for metered work. That's the entry-level salary. Agents WILL climb the same corporate ladder: outcomes next, then a cut of the top line.

Every company has a Brenda: the person who's been there 30 years, knows where everything lives and why, and quietly holds the whole place together. Nobody fires Brenda. That's the prize for not limiting your agent. Let it grow inside a customer long enough and it becomes their Brenda.

Even if customers could export everything she knows, replacing her would be a huge pain in the butt, and a big risk. Software has always had this kind of lock-in. Companies still run on COBOL. But COBOL just sat there. Brenda gets better every year. Whoever becomes Brenda first wins. Everyone else is interviewing for her job.

A Brave New World

None of this is easy. If it were, everyone would be a $1T overnight success. But I think it's the only rational way to build a new startup today. Build the old way, and every time the models get better, your moat gets weaker. Build this way, and every time the models get better, your product gets better too. For free.

It's time to free AI from the shackles of your pricing model.

I write about what I’m seeing in tech and AI. Follow along for the next one.

← All writing