Thoughts on Agentic AI Part 1

Over the last few weeks, I adopted agentic coding. As a seasoned developer with 12+ years of experience, I never expected our field to be disrupted this fast.

When the first AI-assisted coding tools came to the market (e.g. GitHub Copilot), they were useful, but not a gamechanger. It felt like autocomplete gone wild. Sometimes it was extremely helpful; the next day it killed your app with an endless loop.

Then in 2024, I first heard the term "Vibe Coding" and thought it was complete bullshit. How could you trust any LLM to write larger-scale codebases, given the performance I experienced with Copilot? That will never happen.

During 2025, things started to get interesting. Windsurf and Cursor improved the toolchain around LLM-assisted programming to a point where some tasks could be handled faster and with higher quality than a human developer. With the first versions of Claude Code this became even better — but you still had to inspect the code heavily for quirks, security issues, etc. Often it still did not compile or run. And everything seemed to take ages.

2026 brought Opus 4.5 and soon thereafter Opus 4.6. They made answers more precise — and therefore faster and higher quality.

Working with these tools daily is now a genuine pleasure. You can further improve the experience with MCP, Skills, CLAUDE.md instructions and more. And I love that my thoughts get executed and iterated on without having to dig into every API myself.

The hype is annoying

"Follow my AI guide!", "This will increase your productivity by 200%", "You will never write a single line of code." You've probably heard these. They are exactly what they sound like: HYPE. People trying to hook you on LinkedIn, X, or wherever, pushing you to adopt their AI tool. My advice: stay calm, do your work, check for yourself, verify — and if you accidentally jumped on a hype train, take a step back and think.

My productivity increased significantly with AI tooling. But maintaining quality and fitness for users takes just as much attention as before. Building a feature is no longer the bottleneck — staying focused, thinking through quality, and seeing things through to the end are still the hard parts, even when you didn't write the code yourself.

Pick some clear thinkers, follow their work, hold on to your own judgment, and keep improving. The hype will stop bothering you, and you'll get a lot more out of tools like Claude Code.

Parallelization is in an early state

For some time now, all major AI tools have been experimenting with parallelizing workflows — subagents, agent groups, and so on. These are starting to work quite well. Giving agents differently primed prompts in particular seems to increase quality and depth of solutions.

But that is not the parallelization I want to discuss here. I mean things like Aperant (formerly AutoClaude)^[1] or self-written agentic setups: automation that runs overnight and leaves you reviewing finished features in the morning.

We took our first steps into that mode of working and did not get good results. I see two core pitfalls:

1. It cannot ask back — so hallucination multiplies

Most of the code from our first overnight runs did not even compile. AI is still unreliable over long unsupervised sessions. Opus improved reliability significantly for shorter tasks, but left alone for a night, errors do not just add up — they compound. If your tooling and safeguards are not solid, things can go badly wrong.

2. Humans still cannot multitask

When you sit down to review overnight output, you are jumping between features, bugs, implementations, and your regular work all at once. That does not work. Good features require deep thinking, not just a feedback loop. Hard decisions take time. If you are constantly context-switching, that thinking never happens — and your productivity suffers in the long run.

I am deeply skeptical of this kind of workflow. There may be use cases like OpenClaw^[2] — which churns out integrations at volume where internal quality is less critical — where it fits well. But for serious software, I do not see it working reliably anytime soon.

Our future is bright

I deeply love working with these new tools. It has never been easier to build something from scratch, test new ideas, or move users forward. Will my role change because of this? Definitely — but I hope to become the "product engineer" type some people are talking about: technical enough to verify code quality and security, but human-facing enough to build software people love to use and are willing to pay for.