Volume 8

Volume 8 - AI Agents

Volume 8 follows one agent, the retail customer-service agent of Sierra's τ-bench, from a loop around a stateless model to a launch plan with a reliability gate, a price per case and a staged rollout. Section 1 measures its loop, planning, memory and tools on the 1,380 conversations Sierra published; Section 2 rebuilds its return flow in LangChain, LangGraph, CrewAI and AutoGen and weighs them against the vendor agent SDKs; Section 3 adds delegated identity, evaluation by pass^k and defences against prompt injection.