EverMind-AI Unveils Raven, an Open-Source Multi-Agent System Aimed at Self-Improving AI Workflows

A new open-source project called Raven has surfaced on Hacker News, positioning itself as what its creators, EverMind-AI, call a "harness of harnesses" for orchestrating multiple AI agents on complex tasks. The system is designed around the concept of recursive self-improvement, meaning it is built to propose, test and adopt changes to its own operating logic over time.

At its core, Raven acts as a "Host Agent" that coordinates both built-in and third-party AI agents, generating task dependency maps to divide complex jobs among specialists. The project is powered by EverOS, a memory layer that the developers say allows Raven to retain context and experience across sessions, which they describe as essential to its iterative self-improvement process.

Raven ships with four built-in agents: Raven-Research for literature reviews and technical analysis, Raven-Code for software development and debugging, Raven-Design for slide decks and visual assets, and Raven-Oncall for unattended, long-running workflow automation. EverMind-AI says these agents post competitive or leading scores on a range of benchmarks, including SWE-bench, PresentBench and DataAgentBench, though these figures come from the developers themselves and have not been independently verified.

The project's documentation showcases three completed projects it says were driven end-to-end by Raven with minimal human input. In one, Raven reportedly worked autonomously for roughly four days across 42 rounds of planning and testing to build a playable first-person shooter in the Godot 4 game engine, complete with a poster, presentation and website. In another, dubbed "Raven RSI," the system is said to have run 172 machine-learning training experiments across seven rounds without a crash, reducing a model-quality metric by 5.8 percent, while also cutting error in a fluid-dynamics simulation and completing a structural engineering search task. A third showcase describes a full product-launch kit, including a browser game and bilingual marketing materials.

EverMind-AI describes Raven's self-evolution mechanism as a "Curator" that can rewrite an agent's memory, planning, tool access and decision-making logic in successive rounds, installing only changes that pass verification checks. The company notes this Curator feature is experimental and distributed with the source code rather than the packaged release. Raven can also connect to 13 third-party agents, including Claude Code, GitHub Copilot and Qwen Code, through a unified interface.

The software is available for installation on Linux, macOS, Windows and via Docker, and the project explicitly labels itself pre-alpha, warning that interfaces and configuration may change quickly. Beyond EverMind-AI's own materials, there is no independent confirmation of the performance claims or the extent of Raven's autonomy in the showcased projects, and readers should treat the vendor-supplied benchmarks and narratives accordingly.