Raven

What Raven does

Raven calls itself the harness of harnesses, built for recursive self-improvement. As a Host Agent it generates DAGs and orchestrates specialized agents for complex tasks, bringing built-in and third-party agents together in one surface. Four built-in agents cover research (Raven-Research), coding (Raven-Code), visual design (Raven-Design) and unattended workflow automation (Raven-Oncall).

The architecture is meant to let Raven propose changes to how agents plan and act, evaluate them, and adopt those that pass validation. A separate tool, Raven Evolver, consumes Raven as a library and tests candidate harness changes against benchmarks. Memory and context carry across sessions through EverOS. The README presents three showcase projects, including a first-person shooter built in Godot 4, and states that Raven is pre-alpha with interfaces and configuration that may change quickly.

Key features

• Generates DAGs and orchestrates specialized agents
• Built-in Raven-Research, Raven-Code, Raven-Design and Raven-Oncall agents
• Acts as a Host Agent for built-in and third-party agents
• Memory and context across sessions through EverOS
• Raven Evolver evaluates candidate harness changes against benchmarks

When to use it

• Delegate a focused task to one agent or run several in a shared workflow
• Run unattended workflow automation with Raven-Oncall
• Experiment with recursive self-improvement loops on coding or research tasks

Who it is for: Developers experimenting with multi-agent orchestration and self-improving agent harnesses who accept pre-alpha software.

How it fits with Hermes Agent

Tagged hermes and hermes-agent on GitHub and described as a host that brings third-party agents together; the opening part of the README does not spell out how Hermes Agent plugs in.

Note: The README states Raven is pre-alpha and that interfaces and configuration may change quickly.

FAQ