Raven

Why it's ranked here

The case rests on engineering visible in the source, not on the showcase. The agent loop carries measured guards against runaway turns, the local control channels require secret tokens, and the build demands coverage on changed lines. Apache 2.0 licensed and installable as a single command. Against that, the authors call it pre-alpha and warn that interfaces and configuration may change quickly, and the headline results, a game built over four days and a training run improved by 5.8%, are self-reported showcase claims.

What's good

Stuck-loop handling is specific: when one call returns the identical answer eight times in a turn the agent is nudged, at twelve the call is refused, and after three refusals the turn ends. When an editor runs it, commands with outside effects such as pushing code or installing packages ask first, and that rule also reaches delegated sub-agents. The local message channels bind to the loopback address, check a per-boot secret before anything else, and cap frames at 1 MiB. Memory backends are found by reading manifests only, so a broken optional backend cannot stop startup.

Tradeoffs

The surface is very wide for a pre-alpha project: a dozen chat channel adapters, a sandbox, browser control, agent-to-agent protocols, a knowledge store and a tracing viewer all sit in one repository, and the core install already pulls a vector database and a model-routing layer. It needs Python 3.12 or newer. The sandbox extra is unavailable on Windows, and browser driving is opt-in because it downloads its own browser. Some design documents are dated records that describe layouts later replaced, so reading the docs needs care.

How to use it well

Fit it to long, multi-step jobs where splitting work across research, code and design agents pays off, and where you can watch a pre-alpha tool closely. Run it inside your editor if you want each risky command shown for approval. Add only the extras you need, such as one chat channel or the sandbox. Use the separate evolution tool only with a registered benchmark: it validates the setup with a check command and keeps sealed test numbers hidden until you finalize, which cannot be undone.

Observed

License

Apache-2.0 (declared in pyproject.toml and package.json)

Language

Python (requires 3.12 or newer), with TypeScript terminal UI, web UI and WhatsApp bridge

Install surface

Python package with a raven console command; optional extras for chat channels, sandbox, browser and tools

Interfaces

CLI, terminal UI, web UI, ACP editor agent over stdio, JSON-RPC over loopback, MCP and A2A client dependencies

Maturity

README states pre-alpha; pyproject classifier is Development Status Alpha

Platform support

Sandbox extra excluded on Windows; Docker build and compose targets in the Makefile

Testing

pytest suite with branch coverage and a changed-line coverage gate in the Makefile