Raven 0.2.0：能自我改写的 harness 编排层
开源项目 Raven 0.2.0 把 Research、Code、Design、Oncall 与 Claude Code、Codex 放进同一任务图共享记忆，并能重写自身提示与策略。
one AI harness is useful. a harness that can improve the harnesses around it is a different idea. @LongTermMemoryE just shipped Raven 0.2.0, a “harness of harnesses” that puts Research, Code, Design, and Oncall alongside agents like Claude Code and Codex on one task graph with shared memory. the interesting part isn’t just the 0.963 Node F1 orchestration score. it’s the loop. Raven can rewrite its own prompts, policies, strategy code, and playbooks. each sub-harness, including the orchestration layer, can be improved. Oncall keeps jobs running for days, whether that’s a research loop or building a full Godot game. open source. Apache-2.0. and they used Raven to make a lot of the launch itself. that’s probably the part I’d test first.
这条是引用着另一条写的：x.com/longtermmemorye/status/2104745468119146913