This blog post shows my adaptation journey with AI, where I am today (October 2026) and what I’m planning to introduce. What I mention here is what works for me and what I found the most value in. Hopefully it can help others.

A brief history from 2023 until the present.

I started like everyone else: prompting ChatGPT, getting a response and copying and pasting it into my codebase. Like in the StackOverflow era but getting the code from GEN-AI. Then Cursor IDE and Claude Code arrived and it was possible to interact with the agent directly in the local code. But up to this point, I was still just chatting, with no workflows or proper direction; the amount of AI Slop was massive. The supervision and review became another bottleneck.

Today

Verification is the base of everything.

Here is where I spend most of my time. Giving the agent the capability to test its own work was a game changer. I’m not talking only about integration tests; I mean giving it a set of reproducible bash commands to let the agent validate its own work. For instance, in OpenWater, my Zig application to parse dive logs, I gave the agent access to a LIMA VM to be able to test on Linux and reproduce the behavior of the CI locally. If I’m debugging the performance of an app, I give my agent the capability to use Hyperfine. In other words, allow it to use what you use as if you were developing by hand.

Potato Skills

The pstack of Poteto was one of the gems that I found in 2026. It is a set of agent skills that turns familiar engineering practices into procedures the agent can follow across different tasks.

My harness

I mainly use Opencode with the Go plan because it allows me to test different models for a more than fair price. I also switched back to Cursor (this time CLI) because pstack was mainly developed for Cursor.

I dropped Claude Code entirely because they did not add support for AGENTS.md. I did not want to maintain different .md files for different providers. I know that I can link @AGENTS.md from @CLAUDE.md but that’s not the point. I know that they reverted this decision recentely but I was already changed to other harnesses. I still use the Anthropic models but not from their harness.

Local AI Observability

I have a local instance of Langfuse connected to my harness. Even if I use my harnesses in subscription mode, it helps me to study and understand the traces. Being a user of Langfuse led me to find and report a bug in its integration with Claude Code where it was reporting the wrong token count because it ignored provider-reported usage, including cached tokens. I wrote in more detail here.

I have to understand what is happening

I do not buy the idea of not reading or reviewing the code at all. I have, of course, reduced the supervision a lot, but I still need (and want) to understand the architecture and the workflow of the application at a high level. One of the most tangible examples is when I gave the agent the instruction to implement the CI with Nix and I did not even know the basic concepts of Nix. It became a back-and-forth of trial and error until I took a pause to understand the basic concepts of Nix. I realized that the agent decided to use the Nix Sandbox when all I needed was the Nix develop approach: https://x.com/dfrojas89/status/2112521437248458976.

AI Memory

Since I use several harnesses, I found this project: https://github.com/akitaonrails/ai-memory. It allows me to have cross-shared memory and context across harnesses.

What comes next

Cloud Agents

I’m still not at this point, but one of my plans is to try Orbs of AMP or Cursor Cloud Agents.

Local AI

I want to go deeper with vLLM and Ollama. I have tried some things but nothing demonstrable.

What I’m not using anymore after trying it.

Spec-Driven Development and Plans

SDD looked promising; I tried it but then realized it became another system to maintain. Changes made outside the SDD workflow left the specs outdated, and I had to reconcile them with the code. I replaced that workflow with ADRs to preserve decisions and their reasoning. I wrote about this experience here

The agent makes plans based on assumptions rather than evidence. I found that most of the time, it generates a lot of AI slop to account for scenarios that will never occur. I replaced this with clear instructions for the agent to provide evidence for every decision or change.

What I never plan to change

I do not plan to delegate the writing to an AI. I do for very limited things like commit messages or english grammar correction. But my articles, technical proposals, etc I do it by myself, but because the AI does not good well enough, is because I do not want to lose my capacity of express my ideas by myself, writing is one of the most poweful tool that we human have.