My AI Agent - First Week

12 days in. The agent is live, running on Railway, and I talk to it daily via WhatsApp.

The vision

An AI agent that centers all my tools and gives me intelligence and speed. It’s aware of my tasks, habits, goals, and finances. It’s proactive — it doesn’t just answer when I ask, it surfaces what I need to know.

How I’m building this

I’ve been doing spec driven development as much as I can, following the method from Martin Fowler’s SpDD article. The flow looks like this:

1. I outline a roadmap for the project — features I want, somewhat of a timeline (not mandatory)
2. Every time I start working, I choose one feature from the roadmap (or add new ones), building a backlog
3. I detail the feature and write an initial_motivation file
4. I ask the agent to write a requirements file based on the SpDD method
5. I talk with the agent and write an analysis_context — a document about the strategy to achieve the requirements
6. I do this same process for a concrete implementation plan
7. Clear the agent context. Ask the agent to implement the plan
8. Test, revise, iterate when needed, ship to production and test

The whole cycle takes less than 30 minutes. So every day I can get substantial work done in less than 2 hours. Actually. most of the time I’ve spent on setup: infrastructure, integrations, etc.

What’s working

Task management has been the most interesting feature so far. Very fast, very little friction. I use it to get the outline for the day, store things to do, say things I just did. It’s become natural — like a brain dump that actually organizes itself.

Obsidian integration for habits and goals was the right next step. I’ve been tracking sleep, exercise, reading, stretching. The agent surfaces my streaks and monthly goals proactively.

The agent itself has been very fast and reliable. Hasn’t failed yet.

What didn’t work

Finances. I started a cycle on financial management, but quickly understood it was more complicated than I thought. Modeling the finance data wasn’t the next obvious step for the project. It would probably make me frustrated, tired, and unsatisfied. I have to give more brainpower to this, so I decided to abort after a first attempt and go to easier quick wins. That was the right call.

What I’m not happy with

The prompt and intelligence. It’s very repetitive and has a fixed script. Despite that, I have very few tools yet, so it isn’t really a problem right now. But going further I should evolve that.

What’s next

Something that is absolutely useful: enabling the agent to schedule his own next executions, so he can actively monitor and check up on me. That’s going to be my next step.

Pace

This experience has been very enjoyable. Though it has definitely slowed down these last few days — I’ve been busy with work and needed some time off programming. But the rhythm when I’m working is real: less than 2 hours a day, shipping one feature at a time.