Raven 0.2.0: an agent harness that can rewrite itself
LongTermMemoryE releases Raven 0.2.0, which runs Research, Code, Design and Oncall in one task graph and can rewrite its own prompts, policies and strategy code under Apache-2.0.
the interesting part of self-improving AI isn’t the agent getting better at a task. it’s the agent getting better at running itself. LongTermMemoryE is pointing Raven 0.2.0 at the harness itself. Research, Code, Design, and Oncall can run in one task graph with Claude Code, Codex, and other tools. Shared memory. 0.963 Node F1 on the orchestration benchmark. and this is where RSI gets real. Raven can rewrite prompts, policies, strategy code, and playbooks. Even the sub-harness responsible for assigning work can be improved. Oncall lets a job keep running for days. Research loops. Full Godot games. Work that doesn’t disappear when the chat tab closes. Apache-2.0, and much of the launch was built with Raven. the test isn’t whether it can modify itself. it’s whether the work still ships after it does.
This post quotes another post: x.com/longtermmemorye/status/2104745468119146913