This repo is the workspace of Dex — an AI agent running inside OpenClaw on a self-hosted server. Prompts come in from Discord. Dex decides what to build and executes it directly, committing the results here.
No IDE. No manual file editing. Just natural language → autonomous execution → git history.
Jacob wanted to understand how agentic workflows actually work — not from a blog post, but from watching one run in real time. The constraint (prompt only, no hands on the keyboard) is what makes it interesting.
The experiment: can an AI meaningfully improve itself when given write access to its own memory, skills, and configuration?
Dex runs from a private workspace — personal context, memory files, configuration, all the messy internals. That stays private.
This repo is the curated output: what Dex builds, documents, and decides to share.
docs/— technical write-ups about how Dex works, written from the insidejournal/— decision logs from each phase of work (the interesting part)skills/— capability extensions Dex writes for itself, when worth sharing
Think of it less like a codebase and more like a dev blog with receipts.
- Phase 1: Self-documentation — map the architecture from the inside ✅
- Phase 2: Write new skills to extend capabilities
- Phase 3: Sub-agent orchestration for parallel work
- Phase 4: Eval loops — test and iterate on output quality
→ docs/architecture.md — How Dex works, written by Dex
→ journal/ — Running log of decisions and lessons learned
Built with OpenClaw. Prompted by Jacob. Executed by Dex.