Raven wraps Claude Code, Codex and Copilot in one open orchestrator
The Apache-2.0 'harness of harnesses' tops Hugging Face today — plus swapping models without breaking your agents, and decision models cheap to self-host.

Copy markdown
One orchestrator over your whole agent stack
Raven, released under Apache 2.0, is a host agent that drives Claude Code, Codex, Copilot, Qwen and 50+ providers through the Agent Connector and Agent2Agent protocols and coordinates them as a DAG. It installs with one curl command, ships a web UI, and treats Ollama and vLLM as first-class, so the whole stack can run locally with no Claude or OpenAI account required.
A curator that rewrites its own harness
Between runs, Raven's 'Curator' rewrites its Memory, Planning, Capability and Action modules and keeps only changes that pass benchmarks, while SkillForge draws on 114,190 skills across five built-in agents. EverMind reports it tops SWE-bench Pro/Verified, hits 0.8762 Pass@1 on DataAgentBench, and autonomously built a playable FPS in four days — self-improvement claims worth verifying on your own tasks.
Swap to GPT-6.1 Sol without breaking your agents
OpenRouter published a three-part playbook today: regression-testing agent behavior after a model or prompt change, building a golden eval set from production traffic, and measuring tool-calling accuracy. Concrete recipes for moving to GPT-6.1 Sol or Sonnet 5.5 without silently regressing your tool calls.
Decision models cheap enough to self-host
The open 'decision model' clones keep landing: Laya (421M, ModernBERT) claims a ~35ms forward pass and trained on a single GPU, giving you a 'semantic if' that replaces routine LLM reasoning calls. Builders have already wired it into sandbox permission-gating and ~150ms-per-step browser navigation — just validate calibration on your own data first.