Coding agent

CodingAgent is an Agent with a bundle of file, shell and web tools and instructions for working on a code base. Every file and shell operation is scoped to base_dir.

from gwenflow import ChatOpenAI
from gwenflow.agents import CodingAgent

agent = CodingAgent(
    llm=ChatOpenAI(model="gpt-5-mini"),
    base_dir="./my_project",
)

response = agent.run(
    "Add a `greet(name)` function to utils.py that returns 'Hello, <name>!', "
    "then write a pytest test for it and run the test suite."
)
print(response.content)

The agent lists the files, reads the relevant ones, makes targeted edits, runs the tests in the shell and reports back.

The bundled tools

Tool	Does
`LsTool`, `FindTool`, `GrepTool`	list, find by name, search contents
`ReadFileTool`	read a file (the instructions require reading before editing)
`EditFileTool`	replace an exact, unique passage
`WriteFileTool`	create or rewrite a file
`ShellTool`	run tests, linters, git, package managers
`LocalFileWriteTool`	save an artifact (report, dataset) to `target_directory`
`WebsiteReaderTool`	read external documentation

Tools passed with tools=[...] are added to the bundle, not substituted.

Options

Parameter	Default	Meaning
`base_dir`	current directory	the directory the tools work in
`restrict_to_base_dir`	`True`	refuse file paths outside `base_dir`
`target_directory`	`base_dir`	where `LocalFileWriteTool` saves artifacts
`include_shell`	`True`	include `ShellTool`
`include_web_reader`	`True`	include `WebsiteReaderTool`
`include_local_file_writer`	`True`	include `LocalFileWriteTool`

name, description and instructions have defaults you can override, and every other Agent parameter applies.

restrict_to_base_dir confines the file tools, not the shell: ShellTool runs any command with the permissions of your process. Run the agent in a container, or set include_shell=False and pass ShellTool(base_dir=..., requires_approval=True) so each command is reviewed. See Tool approvals.

Long sessions

A coding session reads many files and runs many commands, which fills the context window fast. The default compaction already replaces older reads of the same file by the latest one; for small-window models, see the recommended configuration in Context management. To build the same tools for your own agent, use build_coding_tools(base_dir=...) from gwenflow.agents.coding_agent.

Get started

Concepts

Models

Agents

Multi-agent

Extend

Memory

Flows

Knowledge

Production

Project

The bundled tools

Options

Long sessions

​The bundled tools

​Options

​Long sessions

The bundled tools

Options

Long sessions