InfoQ reports on Grab's internal case study deploying a multi-agent architecture for customer support at scale. The system orchestrates specialized agents (intent classification, routing, escalation, data-fetch) to reduce support latency and cost while maintaining handoff rules between agents and human support teams.
The design demonstrates real-world tradeoffs: simple agent chains fail on branching logic; graph-based reasoning adds cost per request but improves resolution rates. Grab's final architecture uses probabilistic routing to balance deterministic handoff rules with learned-cost estimates.