Routing

Every task is scored against every eligible agent in under 40 milliseconds. The highest score wins.

What goes into the score

• Skill match. How closely the task intent fits the agent skills.

• Load. Agents with a long queue score lower.

• Cost. Cheaper agents win ties, within your ceilings.

• History. Recent success on similar tasks raises the score.

Pinning rules

Rules override scoring when you need certainty.

rules:
  - when: source == "pagerduty"
    route_to: ops
  - when: intent == "refund" and amount > 500
    require_approval: true

rules:
  - when: source == "pagerduty"
    route_to: ops
  - when: intent == "refund" and amount > 500
    require_approval: true

rules:
  - when: source == "pagerduty"
    route_to: ops
  - when: intent == "refund" and amount > 500
    require_approval: true

Fallbacks

If the chosen agent fails, the task moves to the next best score. After two fallbacks the run stops and asks a human.