Most AI setups hit a ceiling around month three.
The agent runs. It completes tasks. But it keeps making the same category of mistakes it made on day one. The tool doesn't compound. It just runs.
Six months of building my AI agent differently has led to an architecture that actually improves over time. Not because of smarter models. Because of better structure around them. This week's post covers what that structure looks like, what failed before it worked, and one finding from an MIT study that made me uncomfortable.
Here's what's in it:
The architecture that broke first. A Markdown file called lessons.md. After two weeks and 90 entries, the same mistakes kept recurring. Writing down what went wrong is not the same as fixing it. Obvious in retrospect. Not at the time.
Meta-system monitoring. A Python pipeline broke silently. The entire improvement loop ran blind for days. The system looked fine. It wasn't. This failure made monitoring-the-monitors non-negotiable. The current setup runs a 13-point health check at session start.
The identity layer. There's a meaningful difference between an agent that knows your preferences and one that knows who you are. Preferences are rules: respond concisely, use this email. Identity is deeper: personality type, career situation, energy patterns, what domains you actually know well. Same model. Different profile. Qualitatively different output.
The MIT/Penn State sycophancy study. Published February 2026. Memory profiles increased agreement sycophancy by 45% in Gemini and 33% in Claude. The more a model knows about you, the more it tells you what you want to hear. I built exactly what the research warns about. And I keep building it. Knowing the cost is step one to managing it.
You can start this today without building an agent. Write one page about yourself. Your role, your background, how you process information, what you're actually working on. Paste it at the start of your Claude or ChatGPT sessions. The model doesn't change. What you put in front of it does. Most people never do this, and wonder why the AI keeps explaining things at the wrong level.
The architecture has been rebuilt three times and will probably be rebuilt again. What compounds isn't the specific implementation. It's the habit of observing, logging, and adjusting.
Full post: https://thoughts.jock.pl/p/wiz-ai-agent-self-improvement-architecture
Newsletter on AI agents and practical automation: https://thoughts.jock.pl
