The Agent Control Plane: Deploying Multi-Agent Enterprise Workflows

About this session

Multi-agent demos are compelling: a planner delegates work, specialized agents collaborate, and complex tasks appear to complete autonomously. Production is where that illusion meets reality. Agents loop, duplicate actions, lose context, select the wrong tools, exceed budgets, and fail in ways traditional software was never designed to handle.

This talk presents a practical architecture for deploying multi-agent systems inside enterprise CRM, ERP, and workflow automation platforms. We will examine how to divide responsibilities among agents, choose between deterministic workflows and dynamic delegation, manage shared state, enforce permissions, and safely execute actions against business systems.

Attendees will learn how orchestration patterns such as routers, supervisors, specialized workers, checkpoints, and human approval gates can make agentic systems observable, testable, and recoverable. The goal is not to create the largest possible swarm of agents. It is to build the smallest coordinated system that can reliably complete valuable work under real-world constraints.

Speaker

Key takeaways

  • Understand when to use a router, supervisor-and-worker model, sequential workflow, or parallel agent execution, and when one agent is enough.
  • Use distributed tracing, structured event logs, replayable executions, versioned prompts, and controlled rollouts to diagnose and improve agent behavior.
  • Separate planning, reasoning, retrieval, validation, and execution responsibilities to reduce ambiguity and improve system reliability.

Related sessions