What I Learned Building AI Agents: From OpenClaw to Claude Code

I was riding the wave like everyone else. Open source AI agents, personalities on Telegram, server integrations, multi-agent orchestration. The promise was real and I bought in hard. I even built something on top of it: Clawdeck, an open source kanban board for managing AI agents and their tasks. A shared board where you and your agents could work side by side. It got 300, maybe 400 stars on GitHub. Thousands of users. I was genuinely proud of it.

Then every morning started the same way. Something broke.

Why Agentic AI Servers Are Harder Than They Look

An update wrecked an agent. The server needed attention. A tool stopped responding. The things you saw on Twitter, agents running autonomously, handling entire workflows, sending Telegram messages, making decisions, were impressive in demo and painful in practice. And the token burn just to keep agents idle on a server was insane. I was spending more time fixing agents than making them useful. The overhead grew faster than the value.

The core problem with agentic AI setups in 2024 wasn’t intelligence. It was maintenance. Every autonomous agent you run is infrastructure you’re now responsible for. Most people building with AI don’t want another server. They want a faster version of themselves.

So I quietly walked away. From OpenClaw, from Clawdeck, from the whole agentic server dream. It wasn’t a decision so much as a slow drift. One missed morning turned into a week, then a month. The fun stopped and I stopped showing up.

That’s the honest story about building AI agents. A lot of people don’t tell it.

What Actually Changed: Context Over Orchestration

Then Claude Code started shipping more. I gave it skills. I gave it instructions. I started building out a proper .claude/ config: a CLAUDE.md that briefed Claude on who I am and how I work, settings that controlled what it could do automatically, agents for specific jobs, hooks that fired in the background. And then I connected it to my Obsidian vault.

Somewhere in there, without a dramatic announcement or a specific moment I can point to, everything changed.

The difference wasn’t capability. It was context, and it compounds.

Every morning I open my terminal and Claude already knows. Not because I briefed it, but because two months of work lives in a shared vault it can read. Notes from client calls, project decisions, what I was building three weeks ago, why I stopped, what changed, what’s next. Tasks open in lst.so, a shared task list where Claude and I work from the same queue. Decisions I made last Tuesday. Ideas I captured at midnight. It’s all there, structured and searchable, and Claude works from it like a colleague who actually reads their notes.

We don’t start from scratch. We just pick up.

How This Works in Practice: AI Agents Across Multiple Projects

I’m working across two client products right now, plus three of my own. Any one of them could come up at any point in the day. When I switch context, I don’t re-explain anything. I don’t paste in background. I just say “let’s work on the recruiting platform” and Claude knows what it is, what we’re building, where we left off, what the open questions are, and what the constraints are. Instantly. That’s not a small thing when you’re operating across multiple projects all day.

The other unlock was having a north star that Claude understands and applies. I have a document called Polaris, a single file that captures what I’m actually optimizing for in life and work right now, my priorities, my non-negotiables, what matters and what doesn’t. Claude reads it. And it starts to show up in how we work together. Not in a corporate-values kind of way. More like, it just knows what I’d say yes to and what I’d push back on, and it factors that in without being asked.

Part of that is lst.so, which I ended up building myself. It’s a shared task list where Claude and I work from the same queue. I create tasks, Claude can read them, update them, log progress, mark things done. No separate agent dashboard, no kanban board living on a server somewhere. Just tasks, shared context, and one place where work actually lives.

This is the thing Clawdeck was trying to solve. I built Clawdeck because I wanted a visual layer between me and my AI agents, a place to see what was running and what needed attention. But the real problem wasn’t visibility. It was that the agents didn’t know enough about my world to be useful. Giving them a board to work from didn’t fix that. Giving Claude a shared task list that lives inside my broader context did.

lst.so replaced the whole setup. Not because it’s a better kanban. Because it’s not a kanban at all. It’s a lightweight task layer that plugs directly into how Claude and I work together every day. Claude checks what’s open, picks up where we left off, logs what it did. I see it all in one place. No server to maintain. No tokens burning while nothing is happening.

That’s what I was trying to build with OpenClaw and Clawdeck. An AI setup that actually knows you, that carries state across sessions, that works with your whole context and not just the last five messages. I just thought I needed a server and an orchestration layer to get there. Turns out I needed a well-structured folder, a vault, and a task list Claude could actually touch.

What This Means for Building AI Agents Today

The concrete setup, if you want to replicate it: a CLAUDE.md per project that gives Claude full context on what you’re building and how you work, a shared Obsidian vault as your knowledge base, per-project context files that capture decisions and open questions, and a north star document Claude references across everything. Hooks handle the background work. Agents handle specialized jobs. The whole thing runs locally, costs nothing to keep idle, and gets smarter as your vault grows.

If someone asks me today whether they should build an agentic AI setup or start with Claude Code and a structured local context, the answer is clear: start local. Persistent context in a well-organized vault outperforms a distributed agent network for individual productivity, at least until agent reliability improves substantially.

Clawdeck was the right idea for a different version of this future. The agentic AI hype isn’t wrong about where things are going. Autonomous agents coordinating across systems, running overnight, handling workflows without you in the loop, that’s coming. But the old rules of building with software are already dead, and the version that changed how I work every day isn’t running on a server. It’s sitting in a terminal to my left, reading from the same vault I write to.

I can’t go back to cold starts. I won’t.