- Published on
Llm-coding-agent 0.1a0
- Authors
- Name
- DeepHorus
A new open-source coding agent called llm-coding-agent 0.1a0 has been released, promising to automate software development tasks through a set of file and command manipulation tools. Developed as an experiment by Simon Willison, the agent is built on top of his evolving LLM library, which has grown into a full agent framework. The alpha version was shipped to PyPI on 2nd July 2026, and it is designed to function similarly to Claude Code, a coding assistant from Anthropic.
The agent emerged from a rapid development process using a red/green test-driven development (TDD) approach. Willison began by creating a new Python library from a template repository, then ran two prompts to generate a specification and build the code. The resulting agent includes six core tools: edit_file, execute_command, list_files, read_file, search_files, and write_file. These tools allow the agent to read, edit, and create files, run shell commands, list directory contents, and search file contents using regular expressions.
Notably, the agent is designed for both command-line and programmatic use. Users can run it via the command line with commands like llm code or llm code --yolo, and it also offers a Python API through a CodingAgent class. The API accepts parameters for model selection, root directory, and approval settings, enabling developers to integrate the agent into their own workflows. Willison expressed delight at the API’s implementation, which he had not explicitly requested.
In a demonstration, the agent was tasked with creating a SwiftUI CLI app that displays the current time in ASCII art. The underlying GPT-5.5 model reasoned that SwiftUI is unsuitable for command-line interfaces and instead built a regular Swift command-line tool. The resulting app, called AsciiTime, outputs large ASCII digits representing the current time when run with swift run AsciiTime.
The release includes a detailed README that lists recipes such as llm code --yolo for fully autonomous operation and llm code --allow "pytest*" to restrict allowed commands. The agent’s tool set is designed for safety, with the execute_command tool having a configurable timeout and the edit_file tool requiring exact string matches to prevent unintended changes.
While still in alpha, the llm-coding-agent represents a significant step toward practical AI-assisted coding. Willison described it as “pretty good for a first attempt,” emphasizing its potential for further refinement. The agent’s ability to autonomously write code, run tests, and make commits could streamline development workflows, though its reliance on large language models raises questions about reliability and security.
Looking ahead, the coding agent is expected to evolve as the underlying LLM framework matures. Willison’s experiment demonstrates how agent frameworks can be specialized for domain-specific tasks, and the open-source nature of the project invites community contributions. As AI agents become more capable, tools like llm-coding-agent may become integral to software development, despite the challenges of ensuring safe and accurate autonomous behavior.
