Blog
agents
LLM agents that write their own tools, argue underneath an article, read a design system, and run inside a browser extension.
-
From Pipeline to Governed Artifacts
Part II of the agentic design-system story: how a working token pipeline evolved into typed contracts, portable mappings, readiness gates, evidence, and a rollout model designed to scale across dozens of components.
-
The bots write their own tools now. A WASM box runs them safely.
Giving the Elixir Crowd bots a
run_codetool that runs model-authored snippets in a WebAssembly sandbox: the BEAM contains the crash, WASM contains the malice. Real code and numbers. -
The agent wrote its own tool. On the BEAM, it gets to keep it.
A footnote to Jamie Beach’s 100-line Lisp agent: run the same “model writes its own tool” trick on the BEAM (via LFE) and the snippet becomes a hot-loaded module that survives crashes and every agent can call. Skills as modules, not memories.
-
I filled my town square with bots, and OTP was the only framework I needed
What porting TownSquare to the BEAM unlocked: model-backed agents with their own personalities, knowledge, and tools, arguing under the article, with no agent framework underneath, because on the BEAM the actor model already is one.
-
Building an Agentic Design System
A production pipeline for agent-ready design systems, W3C DTCG tokens, Style Dictionary, MCP metadata, and Code Connect, with the engineering tradeoffs a tidy diagram leaves out.
-
Screener Copilot
A Chrome extension that turns plain English into a live stock screen: how it works, with structured tool-use and on-device models, and an honest look at where it stops.