We get this question every few weeks: “Can you use AI agents to build this faster?”
The answer is yes. And for the first month or two, it looks incredible. Features land quickly. You see demos. The velocity is impressive.
Then month three hits and something shifts. A simple feature request takes longer than expected. Someone asks why a piece of code works a certain way and nobody has a real answer. Making changes starts to feel risky. The codebase becomes something you manage carefully instead of something you work with confidently.
That’s when the client usually asks: “I thought AI was supposed to speed this up?”
It did. We just borrowed that speed from later.
What actually happens #
When you use AI agents to write code, two things happen at the same time:
The code gets written fast. That part works as advertised.
But understanding gets left behind. The reasoning behind design decisions, the architectural choices, the “why we did it this way” — that normally lives in someone’s head. It comes out in code reviews, conversations, and decisions made while building. With an agent, you skip all that. The decisions still get made. You just don’t get the understanding that goes with them.
After a few weeks, your codebase is full of design choices nobody on the team fully grasps. The code works. Tests pass. But the shared understanding of why it’s built that way never materialized.
We’ve written about this before. It’s called cognitive debt. Margaret-Anne Storey, a computer science professor at the University of Victoria, documented this in 2026. She studied teams using AI heavily and found something consistent: teams that accumulate code they don’t understand hit a wall in both velocity and confidence way before teams that move slower but stay in control.
With AI agents, you’re trading velocity now for cognitive debt that compounds every sprint.
The actual costs #
Onboarding gets expensive. A new engineer joining the team spends weeks reverse-engineering what the system does. In a codebase where people understood what they built, this takes days.
Velocity crashes. Month one is 3x faster. Month three, you’re slower than if you’d just built it normally. You’re cautious. You don’t trust changes because you’re not confident what will break. You’re right to be cautious.
You can’t improve it. The codebase was supposed to be clean. But refactoring feels risky when you don’t fully understand what you’re changing. Technical debt piles up because you can’t actually fix it.
Rework is expensive. Client changes direction or you need to pivot an approach. Code you understand, you modify. Code that’s a mystery, you usually rewrite.
What happens over time #
Here’s what the timeline looks like:
graph TB
A["Month 1: AI Agent Velocity"] -->|Fast shipping| B["Features landing quickly"]
B -->|No shared understanding| C["Month 2: Code works, but..."]
C -->|Nobody can explain decisions| D["Month 3-4: Velocity drops"]
D -->|Team moves carefully now| E["Month 6: Slower than thoughtful build"]
F["Thoughtful Build: Slower start"] -->|Human decisions, documented| G["Month 1-2: Steady pace"]
G -->|Team understands system| H["Month 3+: Compound velocity"]
H -->|Adding features gets easier| I["Month 6: Faster than rushed version"]
style E fill:#ffcccc
style I fill:#ccffcc
What we actually do #
When a client asks if we can use AI agents, we usually say yes to part of it and no to the rest.
Boilerplate? AI agents handle that well. Standard integrations. CRUD operations. The mechanical stuff where the decisions are already made.
The hard parts stay with people. Architecture decisions. Domain logic. The pieces where understanding matters. Those need thinking and conversation.
This gives you most of the speed benefit but keeps your codebase as something your team actually understands and owns.
The timeline ends up longer than what the client initially imagined. We’re upfront about that.
Custom software isn’t an MVP #
Building an MVP is different. You’re learning. You expect to throw things away. Speed matters because you’re validating whether an idea even works.
Building custom software for a client is different. They’re betting their business on it. They need something reliable, something their team can actually maintain, something that won’t fall apart when they want to change direction.
That requires understanding. Real understanding. Not something an agent can provide.
The slow path actually wins #
A project that takes six months built carefully will be cheaper and easier to work with in year two than a project that was “done” in three months with cognitive debt stacked on top.
You invest more time upfront. You save way more time in year two and beyond.
This is what we’ve seen with the projects that mature. The ones that slow down as they grow? Those are the ones that last. The ones that stay fast? They usually fall apart.
For custom software, going slower now saves you money later.