Routing tasks between agents based on live execution state

Rather than statically assigning one agent to a job, systems are emerging that inspect live execution evidence, permission scope, and remaining budget to decide whether a task should stay with the current agent, be handed to a specialist, or be escalated to a coordinating layer. This turns multi-agent orchestration into a routing problem solved at runtime instead of a fixed pipeline configured up front. Developers wire this in front of existing agents like Claude Code, Cursor, or Codex as an orchestration layer.

Sources

• coco-research/coco - GitHub
• wang2122/sprix-sage-router - GitHub