Agents, Graphs, Loops¶
Run Claude Code and Codex agents as workflows you write in Python. AGL can use your subscription limits so API keys are optional.
You write a workflow as an async Python function. Each step runs an
agent role on a git branch AGL makes for the run, so
your checkout stays as it is. When the run ends, its work waits for you on one branch,
agl/<label>.
Features¶
- Resume where you stopped. Every step is recorded, so a crashed or stopped run picks up where it left off, and finished steps don't run their agents again.
- Agents never touch your checkout. Each run and each worktree works on a branch of its own.
- Parallel agents land only when the build passes.
- Claude Code and Codex in one workflow. Pick the model and effort per role.
- Typed in, typed out. Pass dataclasses into prompts and get dataclasses back.
- Questions mid-step. Any agent can ask a question in the terminal and wait for the answer in the same session.
Getting started¶
Before a run starts or resumes, AGL checks that each tool the workflow's roles use is ready. If one isn't, AGL refuses the run before any agent spends tokens.
Requirements¶
-
uv. If you don't have Python 3.14 or later, uv installs it for AGL.
-
git, with your name and email set. AGL commits agents' work under your name.
-
Claude Code, Codex, or both. Either one is enough if your workflows only use its models.
Claude Code¶
AGL runs Claude models through Claude Code. Log in once and AGL uses that login.
Then follow the login steps. If claude isn't found,
install Claude Code first.
You can also set ANTHROPIC_API_KEY instead. When it is set, AGL uses the key, not your login.
Codex¶
AGL runs OpenAI models through Codex. Install it and log in.
You can also log in with an API key:
Check the install¶
Check that AGL is installed:
If your shell doesn't find agl, run uv tool update-shell and open a new terminal.
Next¶
- Run workflows - run a workflow on your project.
- Download workflows - get workflows other people published.
- Build workflows - write your own.