Raven: an AI agent manager that builds games and upgrades itself

Raven is an open-source meta-orchestrator for AI agents, designed to act as a general contractor for complex engineering tasks. Instead of doing all the low-level work directly, it breaks goals down into directed acyclic graphs and coordinates both built-in and third-party agents to deliver end-to-end projects.

Its headline demonstration is a four-day autonomous run. Driven only by an initial human brief, Raven executed 42 rounds of planning, implementation, and testing to build a playable first-person shooter in Godot 4. Along with the arena boss-fight game, it produced the project website, promotional posters, and slide deck without human intervention.

What it can do

Raven handles end-to-end execution across software engineering, research, and scientific computing. In machine learning experimentation, it managed pre-training runs for nanochat over seven rounds and 172 individual executions without crashing once, shaving 5.8% off val_bpb under a strict 20-minute single-GPU budget. In fluid dynamics, it reduced overshoot in dam-break simulations by three orders of magnitude. The project even generated its own launch materials, including its README, slides, posters, and a browser-based mini-game.

Out of the box, Raven includes four specialized subagents: Raven-Research for deep technical analysis, Raven-Code for software development, Raven-Design for layouts and slide generation, and Raven-Oncall for unattended background runs. It also provides presets to hook into 13 third-party agents such as Claude Code, Codex, and Kimi Code over ACP, CLI, or OpenAI-compatible APIs.

How the self-improving harness works

Instead of attempting the risky task of modifying base model weights, Raven targets recursive self-improvement at the harness level. It decouples the agent loop into four modular strategies: what memory is visible, how planning proceeds, which tools are exposed, and how actions are vetted before execution.

An experimental module called the Curator iteratively rewrites these four components. Crucially, the benchmark evaluation criteria are locked and immutable, meaning the AI cannot alter the scoring rubric to inflate results. Candidate harness improvements must measurably beat the previous baseline on standardized tests before they are adopted. Backed by the EverOS memory runtime and access to over 114,000 indexed skills, successful procedural refinements persist across future sessions.

How to run it

Getting started requires minimal setup. The repository provides one-line shell scripts for Linux, macOS, WSL2, and native Windows PowerShell. Users who prefer isolated environments can spin up the stack with Docker Compose and access the dashboard at localhost port 18793.

Once installed, running raven web launches the browser-based workspace. From this dashboard, users can track task DAGs, configure connected subagents, inspect persistent memories, and manage generated deliverables directly.

Why it matters

We think Raven takes the right pragmatic path toward self-improving AI by focusing on the operational harness rather than fine-tuning weights. Developers managing overnight code refactoring, iterative data science tasks, or complex asset creation will find it a compelling orchestrator to test today.

It also highlights an industry shift: as agents become capable of tuning their own execution loops, prompts and hardcoded workflows lose their defensibility. The real leverage shifts to defining strict, tamper-proof evaluation standards that keep autonomous agents aligned with real-world deliverables.