Luis Paredes Franco
← Projects

42 tests passing · 7 skills it wrote itself

A coding agent built from scratch

Builder, with the AI Guild · INIT FIU, AI Guild · Oct 2026

A minimal coding agent in Python, built with the INIT AI Guild at FIU to see what really happens inside tools like Claude Code and Cursor. Then we had it build its own terminal UI and write new skills for itself.

Code ↗LinkedIn post ↗

1Why build one

I use coding agents every day, and I wanted to understand them from the inside instead of treating them as magic. At the INIT AI Guild we built our own, small enough to read in one sitting.

2What is inside

The whole agent is one loop: send the conversation and the tool definitions to a model, run whatever tools it asks for, send back the results, and repeat until it answers without calling a tool.

  • Tools. Run shell commands, read files, write files, and edit by exact string replacement.
  • Planning. A TODO list the agent creates and updates during multi-step tasks.
  • Skills. Short Markdown instructions the agent discovers and loads only when a task calls for them.
  • A subagent. A read-only explorer for focused research, so the main conversation stays small.
  • Compaction. Long conversations are summarised automatically once the prompt grows past a threshold.
  • Cost tracking. Tokens and cost are reported after every turn.

3Then it worked on itself

Once the loop was working, I pointed the agent at its own repository. I asked it to build its own terminal interface, and it did, using Textual. I asked it to add more skills, and it wrote seven new ones, updated the README and added tests. All 42 tests passed. The run that wrote the skills cost about half a cent.

4What I took from it

The loop is simple. Most of what makes a coding agent good sits around it: clear tool descriptions, a plan it keeps up to date, instructions loaded only when needed, and a context that stays small. Those are product decisions as much as engineering ones, and knowing them helps me judge what is realistic to ask of an agent at work.

5Limits

This is an educational harness. It has no sandbox, no approval step before commands and no path restrictions, so it should only run in a project under version control.

Stack: Python · OpenRouter · Textual · Markdown skills