Intel / blog

A Coding Agent from Scratch

Source: blog.softwarewrighter.com
Author: softwarewright
Date: Published Sep 16, 2026

Context

The educational coding mlplcode makes completion a condition checked by software. Its model proposes an action, the program parses and authorizes it, executes it, and feeds the result into the next turn. The implementation article describes a verifier that accepts completion only after a file has changed and a test run has passed after the latest edit.

If the model announces success too early, it receives a failed-verification observation and continues. This small example makes a central agent-design decision visible: the model chooses what to try, while the surrounding program checks whether the required evidence of completion exists.

Key takeaways · AI-distilled

The agent is implemented using a custom Rust-based ML domain-specific language the author calls sw-MLPL, rather than wrapping an existing agent framework, and is released MIT-licensed with an interactive playground for the language itself.
The implementation and its tests are documented as an Emacs org-mode literate-programming document, so the code and its explanation live in the same file rather than as separate prose docs.
The agent currently only runs locally against Ollama with a coding LLM and does not yet support connecting to cloud LLM providers.
The author says the project was inspired by OpenCode's approach but is not a port of it — it's an independent implementation built on the author's own DSL.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source blog.softwarewrighter.com