State of Agentic AI in April 2025

ByPeter Volfon2025-04-30

People are posting examples of how agentic AI creates, refactors, and restyles entire applications for them, hyping up the technology.

I finally took the time to give agent mode a go. Admittedly, I did not do it in the "conventional" way, i.e. using Cursor, Windsurf, or an officially supported VSCode installation with Copilot. I'm a Linux user (Fedora to be more specific), so I work with the unofficial VSCode flatpak distribution, and the open source Continue extension. But I don't think it had too much of an effect.

Configuring the Continue extension and setting up the models is very straightforward and well-documented. Normally I would use Codestral from Mistral, but it doesn't support agent mode yet, so for this test, I used Gemini instead.

I also configured some MCP servers to make all the necessary tools and the latest docs of every relevant library available to the agent.

The latter step is necessary if you're working with recently released versions of widely used libraries, or relatively niche ones. By recent, I mean (major) versions released after the cut-off date of the foundation model's training data.

After setting up everything, I gave the agent a very simple task: restyle a simple Next.js app using TailwindCSS v4 and DaisyUI v5, the current versions of these libraries.

It failed badly, getting totally confused about library versions or trying to apply various patterns from the old ones. No explanation helped. Maybe Cursor or Windsurf would have worked better, but I prefer to stay away from these products, although Windsurf seems to be quite privacy-respecting. I'm curious to see how Zed's agent mode fares when it's publicly released in May.

After using various AI coding assistants for some time, and having worked with developers and engineers of all levels who also use them, these are the pros and cons I've personally observed:

  • Coding assistants are okay for small auto-completes, copy-pasting boilerplate, or implementing well-know algorithms.
  • Writing documentation is much faster with AI tools. There can be no excuse anymore for not writing docs.
  • Next edit prediction is a great feature.
  • LLMs are decent coders, better than the typical developer in many tasks, but at the same time they are also terrible engineers.
  • They can actively prevent developers from improving their hard-skills, and can reduce critical thinking skills.
  • The improved "productivity" of "it works developers" puts extra workload on code owners and reviewers.

Like other advanced technologies, current AI tools are a multiplier not only for competence, but also for incompetence.