Building AI Agents
11.05.2026 08:16 · 👁 3.4K
Why do most AI products fail while Claude/GPT keeps winning?
They nailed personal assistant and code companion. The reason isn't model quality, plenty of wrappers run the same underlying model. Claude was just first to hand users the AI and the instruments to use it, without assuming what you'd do with them. You bring the use case, it brings the tools and reasoning.
Most wrappers fail because they try a different move: invent a workflow, wrap an AI around it, sell it. But workflows are niche. They fit one person's mental model and break for everyone else's. Users have a lot of friction trying to use them, because they must adapt.
There's a third option almost nobody's building:
Pick a domain. Give the AI the instruments a junior employee in that domain would actually use. Don't define workflows, let behavior emerge from the AI working alongside the user. Make it accessible without technical skills.
Think about it like hiring someone new. A salon owner doesn't hand their new receptionist a 40-page manual. They sit next to them for a week, correct them, show them how things are done here. "No, we don't say that to regulars." "Yes, always offer the upgrade when person buys." etc. The new hire has the tools already, phone, calendar, booking system. What they're learning is how this business uses them.
That's what an AI agent should be. Not a pre-built workflow. A junior employee with the right instruments on their desk, learning from the person who runs the business.
If you're building AI products, your customer doesn't care about prompts or workflows. They need a result. A booking confirmed, a lead processed, a question answered at 11pm. Actual value, not another iteration. And you can't encode that value from the outside, because the knowledge of how to answer, when to offer an upgrade, how to politely say a slot is taken, that lives in the owner's head.
Constrain the tools, not the behavior. A salon needs WhatsApp, Instagram, Telegram, a calendar, notes on regulars. Those aren't workflows. They're instruments. Set up the desk, let the owner teach the AI the same way they'd teach a new hire.
I ran into this exact tension at work recently. A product guy said: "Our users need to see what's important in their inbox. Let's flag every message with sales intent as important.", I pushed back: "are you sure that's what all of them want? Why not give the AI the ability to flag, but let it figure out how by interacting with each user? Different people prioritize differently. There's no catch-all."
We built the capability without hardcoding what "important" means. The "how" gets discovered per user. Same principle.
This is where I think the real value of AI lives. Not in bigger models. In putting capability into environments where emergent value can happen, for the use-cases that have been skipped over.
But here's where it breaks down. Between "I know what I want" and "AI does it for me" stands the integration wall. Connecting to WhatsApp, Instagram, a booking system, you need code, edge case handling, deployment, monitoring. Every step is technical. Every step requires skills the business owner doesn't have.
I'm building in that gap: agents for real businesses, taught in plain language, connected to the tools they already use. I'll share what worked, what broke, and what I'm still figuring out.
