The question you're probably asking
You've seen the demos. An agent like Claude Code reads a project, writes the files, runs the tests and fixes its own mistakes. An admin screen that used to take two days comes out in ten minutes.
Then you open an agency quote: forty days, three profiles, a project manager. And you wonder what exactly you're paying for.
Maybe you've also had the opposite experience. You ran the agent yourself, it worked great, until the day a payment didn't go through and nobody knew why.
My answer fits in one sentence: AI agents don't replace agencies. They replace the way agencies used to bill. And that changes a lot for you.
What you were really paying for
Think of a house renovation. On the invoice, you pay for the builders' hours, a site manager, an architect and insurance.
An agency quote works the same way:
- the builders: the developers writing the code
- the site manager: coordination, meetings, follow-up
- the architect: technical choices and product trade-offs
- the insurance: someone who answers when things break
Historically, most of the billed days went to the builders. Writing code was the slow part, so it was the expensive part. Everything else was buried inside it.
Agents went after exactly that line.
| Before agents | With agents | |
|---|---|---|
| Writing the code | Most of the quote | A fraction of the time |
| Coordinating a team | Expensive, because the team is big | Reduced, because the team shrinks |
| Making the right choices | Buried in the quote | Becomes the real value |
| Answering when it breaks | Included, in theory | Still needed, never automatic |
What agents make obsolete
Padded quotes on standard work. An admin panel, a login page, a CSV export: these building blocks exist in thousands of projects. A well-driven agent produces them fast. Billing them as craftsmanship gets hard to justify.
The pyramid. The classic agency model is a senior in the sales meeting and juniors on the code. With agents, a senior directly produces what several juniors used to, with their own eyes on every line.
Coordination that costs more than the work. Four people on a project means meetings, handovers and misunderstandings. Fewer people, less friction.
Three-week round trips. You ask for a change, you wait for the next sprint. When execution takes an hour, waiting three weeks no longer makes sense.
To be fair: good agencies have understood this and are adapting. What's disappearing is a business model, not necessarily the companies that ran it.
What agents don't do
Choose. An agent builds what you ask for, with the tools it knows best. It won't ask whether that's the right stack for your budget (your AI may have already picked it for you), or whether the feature deserves to exist (the right feature beats building them all).
Say no. An agent says yes to everything. Handy for moving fast, dangerous for a product. Part of what you pay a good provider for is the requests they talk you out of.
Review its work with fresh eyes. An agent writes code that passes its own tests. When I audit apps built with AI, the problems are almost never where you'd expect. AI often nails the hard part and misses the obvious one: a forgotten access rule, a payment counted twice.
Take responsibility. If your app goes down at 11pm on a Sunday, an agent has no phone number. You need someone who gets the alert, understands what's happening and knows how to roll back.
Know your business over time. An agent starts almost from scratch every session. An instructions file like CLAUDE.md helps a lot, but it doesn't replace someone who knows your product, your customers and the history of your decisions.
What my prospects taught me
In spring 2026, I was mostly offering coaching to founders who had built their app with AI: helping them use their tools better, secure the sensitive parts, level up.
Between April and July, nobody wanted that format. The people who contacted me all wanted the same thing: someone to hand their product over to, now that it had users.
It's not that they couldn't learn. They had better things to do. Their time was worth more in customer meetings than debugging a Stripe webhook.
I reshaped my offer accordingly. And I take a lesson from it that goes beyond my case: agents didn't remove the need to delegate. They changed who you delegate to, and at what price.
The new model: a senior and their agents
The model emerging looks like this: an experienced developer who makes the choices, drives agents for execution, reviews what they produce and stays accountable for the result.
This model makes something possible that custom development rarely offered: a fixed price. When execution is fast and predictable, uncertainty concentrates on decisions. A senior can frame them upfront and commit to a flat fee.
Conversely, ask yourself this when you see a day-rate quote. If the provider uses agents and keeps billing the same number of days, who keeps the time saved?
This model has its limits too, and you should know them:
- One person is a single point of failure. Require the code to live in a repository you own, documented, so another developer can pick it up.
- It has a maximum size. A product with several teams, integrated design and heavy regulatory constraints still needs a real organisation.
- Review isn't optional. A senior who lets the agent do everything without reviewing is vibe coding billed at expert rates.
5 questions to ask your next provider
| Question | A good answer | A red flag |
|---|---|---|
| How do you use AI? | For execution, with human choices and review | "Not at all" or "it does everything" |
| Who reviews the code, and how? | Systematic review, tests on critical flows | No clear answer |
| Who owns the repository, the accounts and the domain? | Your company, from day one | The provider, "to keep things simple" |
| Fixed price or day rate? | A written scope, a price, and what happens if the scope changes | Estimated days with no scope |
| What happens after delivery? | Who gets alerts, who fixes, how fast | "We'll see" |
The third question matters more than it looks. If you ever need to switch providers, getting back what's yours should be a formality.
So: you, a freelancer or an agency?
You with an agent, if you're validating an idea, building an internal tool with no sensitive data, or you want to learn. It's the best time in history to do it. The Claude Code guide gives you the basics.
A senior working with agents, if your product has paying customers, reliability matters and your budget can't cover a full team.
An agency, if your project needs several disciplines in parallel, several teams, or a contractual framework only a larger organisation can provide.
The wrong choice is paying 2020 prices in 2026. Or carrying a product your business depends on alone, because an agent makes you feel like you have a team.
Want to hand your product to someone who works this way?
I build and take over custom products at a fixed price, with agents for execution and my senior developer judgment for choices and review. See the offer or let's talk about your project. If your app already exists, a code audit is a good place to start.
For more articles on building products with AI, subscribe to the newsletter.
