Paintings Subscribe Tiếng Việt

22 November 2026

Deep Dive/Moving faster than we can understand

A year and a half of working with autonomous agents: AI can make development 5–8 times faster, but the bottleneck moves from writing code to understanding, controlling and steering it.

I have worked with autonomous agents for about a year and a half. Beyond the agents I use at work, I built my own to run personal workflows almost around the clock.

After a while, I noticed something few people say plainly: AI makes us faster, but it does not make us understand faster.

Five to eight times faster, but faster at what?

With autonomous agents, development speed can rise five to eight times. “AI-first” thinking pushes expectations even further than lean startup ever did.

Do the maths: a team of eight, each working eight times faster, produces the output of a team of sixty-four.

But that is only the speed of writing. What about the speed of reviewing, testing and understanding? Who guarantees the overall quality of eight times as much code?

When the system outgrows the people

End-to-end automated testing helps for a while. But once a project is complex enough, spanning many domains with tangled dependencies, the limit is no longer speed. It is human comprehension.

Multi-agent systems, sold as the answer, can become the source of losing control. Agents write code, fix bugs, optimise and call each other. A small change ripples outward, and nobody truly understands what is happening end to end.

We lose control not for lack of tools, but because we have passed the limit of what we can understand.

Where AI really shines

To be fair, for familiar, well-bounded problems, such as a blog or a customer system with a known process, AI and agents work brilliantly. The problem is already framed, people understand the requirements, and going faster does not break control.

The trouble starts when we carry the same approach into large systems we do not yet understand.

The cost of going too fast

When everything is pushed to be faster, the downside appears:

  • Speed skips deep understanding. The code runs, but nobody really understands the system.
  • Quality is replaced by “good enough”. Technical debt piles up quietly until it cannot be controlled.
  • Automation creates a false sense of control, while people drift further from the system.
  • People are pushed into operating instead of thinking, and over time their judgement weakens.
  • Productivity rises but value does not, and complexity grows for nothing.

And one obvious consequence that rarely gets said: when there is no time left to verify, testing moves outside the team. Users become the testers. Production becomes the test lab. Every fast release is a bet.

The real question

In the end the question is not whether AI makes us faster. It clearly does.

The question is: are we moving faster than our own ability to understand and control what we build?

The things that do not scale like code, such as judging right and wrong in a complex context, seeing long-term consequences, and deciding what to build and what not to build, are still human work. And they are becoming the most valuable part of the job.

For fellow engineers, I think the most important skill of the next few years is not writing better prompts. It is keeping our ability to understand deeply, even when machines write most of the code.