llm-coding-agent 0.1a0: GPT-5.5 coding agent and tools

Simon Willison released llm-coding-agent 0.1a0 on 2nd July 2026, a slop-alpha Python package on PyPI that implements a Claude Code style coding agent built on his evolving llm library and agent framework. The package can be run via uvx and exposes a suite of file, search and shell tools plus a Python CodingAgent API.

What is llm-coding-agent 0.1a0?

llm-coding-agent 0.1a0 is a first-attempt coding agent library published as a slop-alpha on PyPI on 2nd July 2026, built on Simon Willison's llm agent framework and designed to read and edit files and execute commands. It was started from a Python library template and developed with a spec, tests and a Fable-authored README that lists recipes like llm code --yolo and llm code --allow "pytest*" --allow "[git diff](/glossary/diff)*".

The release includes a command-line usage demonstrated as uvx --prerelease=allow --with llm-coding-agent llm code and a Python API centered on a CodingAgent class, shown as CodingAgent(model="[gpt-5.5](/glossary/gpt-5-5)", root="/path", approve=True).run("Fix the failing test in tests/test_parser.py") in the README.

How do its tools and API work?

The package exposes a named set of tools for file and shell work, each with explicit behaviours: edit, read, write, list, search and execute. CodingTools_edit_file replaces an exact string and returns a diff; CodingTools_execute_command runs a shell command and returns combined stdout and stderr plus an Exit code line, with a default timeout of 120 seconds and a maximum timeout of 600 seconds; CodingTools_list_files returns at most 200 paths and skips hidden directories, node_modules, pycache and anything in.gitignore for git repositories.

Other tool specifics come from the README and spec: CodingTools_read_file returns numbered lines like cat -n, with an offset parameter (0-based first line) and a limit parameter (default max 2000 lines) for paging; CodingTools_search_files returns matches in the format path:line_number:line, capped at max_results (default 100) and accepts a glob filter; CodingTools_write_file creates or overwrites files and creates parent directories as needed. The README shows these tools listed via uvx... llm tools and documents the exposed signatures.

The CodingAgent API shown in the release binds a model name (example gpt-5.5), a root path and an approval flag, and the agent runs by issuing file and command tool calls. Willison notes occasional manual testing using an OpenAI API key in the environment while developing the project.

The author exercised the agent interactively: running llm code --yolo and prompting it to scaffold a small demo project produced a Swift-based command-line program. The agent’s reasoning reportedly observed that "SwiftUI isn't suitable for a true CLI" before producing a runnable ASCII-time example.

Why it matters

This release demonstrates a compact, tool-focused approach to coding agents: packaged tools with clearly documented edge cases and limits (for example, a 600 second maximum command timeout, a 200-path list_files cap, and a 2000-line read_file limit). That level of explicit tooling makes it easier to audit and verify what an automated coding agent can change in a repository, and it provides a concrete example of integrating an LLM-driven agent into a developer workflow via both CLI and a Python API.

For developers experimenting with agent architectures, the combination of a published PyPI slop-alpha, a visible spec and test-driven commits offers a replicable blueprint rather than an opaque demo.

What to watch

Watch the project repository and the PyPI package for updates beyond the 0.1a0 slop-alpha, and check the README's recipe list and llm tools output to see how tool capabilities evolve. The next meaningful signals will be expanded tool behaviours (for example larger read limits or different file-safety checks) and any post-alpha changes to the default model binding shown as gpt-5.5 in the API example.