Argonauts Code：以代码定义工作流的开源多智能体编排器

Disclosure up front: I'm the author. It's Apache-2.0 and free, with no hosted part and no pro tier.

The problem I kept running into with multi-agent setups was that the process lived in a prompt. Plan, build, review, run the tests, merge: the agent did those when it remembered to. So in Argonauts Code the workflow is code. A workflow is a TOML file of steps (or a script), the built-in ones are written the same way, and the runner executes every step every time. Agents just fill the roles.

How it's put together:

• Each CLI is driven through its headless structured output, not by scraping a terminal: claude -p --output-format stream-json, codex exec --json, opencode run --format json. ACP agents and a small manifest cover other CLIs.
• Builders work in separate git worktrees. Checks run on each task and again on the exact commit you'd merge.
• If you don't name reviewers, review goes to a different provider than the builder where the crew has one.
• In the Council workflow, seats answer without knowing which model the others run, then talk it out until they agree on a plan.
• Local models write code through OpenCode. For reviews, judging and summaries, local servers (Ollama, LM Studio, llama-server, vLLM) are found automatically, and any other OpenAI-compatible server takes a few lines of config.
• Headless use: argonauts run --crew --plain --limit-usd 2 "...", an append-only log of every step, argonauts run-review to compare a run's predictions with what happened, and argonauts mcp serve.

Testing is where I spent the most thought. None of the 10k+ tests call a model. Each CLI's output is recorded once against a specific CLI version and replayed, and fake CLIs play scripted scenarios for everything else. When a CLI updates, a rerecord script makes a few live calls on its cheapest model and reports which recordings changed shape.

On rule 7: most of the code was written by Claude Code agents running in parallel lanes, with separate reviewer agents. I designed it, approved each screen as a text drawing before it was built, and decided what merged. Every change goes through a PR with CI (clippy with warnings as errors, cargo deny, the full test suite).

What I don't have yet is a real measurement of whether cross-provider review catches more real bugs than it adds noise. A fair test needs a lot of tasks, repeats and a single-agent baseline, and I haven't run one. If you've done something like that, I'd like to hear how you set it up.

It's alpha, macOS and Linux only. The video is a scripted tour on sample data, and its run window, galaxy-style Behaviors view and the mascot arriving at session start come in the next version.

Repo: https://github.com/sheng-tse/argonauts-code

The built-in workflow files, if you want to see the format: https://github.com/sheng-tse/argonauts-code/tree/main/crates/core/templates/workflows