CORAL debuts autonomous multi-agent evolution

This reads less like another agent demo and more like a serious control plane for long-running agent societies. The interesting part is not raw autonomy, but the machinery around it: memory, interruption, isolation, and safe evaluation.

• –Isolated workspaces plus separated evaluation directly target reward hacking and contamination, which are common failure modes in multi-agent loops
• –Shared knowledge layers for attempts, notes, and skills make the system behave more like an accumulating research org than a stateless agent swarm
• –Heartbeat-based reflection is a practical fix for agent myopia: it forces consolidation instead of endless local thrashing
• –The reported 3-10x improvement rates and 10-task SOTA gains suggest the architecture matters, but the results will still depend heavily on task quality and grader design
• –Native integration with Claude Code, OpenCode, and Codex positions CORAL as orchestration infrastructure, not a model alternative