What's changed: Added per-section figures (cert-figure-retrofit). New AI-103 Chapter 3 (agent implementation = Foundry Agent Service, composition = model + instructions + knowledge + tools, tool types = function calling/file search/code interpreter, threads and runs · multistep reasoning, multi-agent = orchestrator/handoff/parallel, safety = least privilege/human-in-the-loop/monitoring & evaluation/Content Safety)
3.2Multi-Agent Orchestration
Understand multi-agent design—splitting complex work across specialized agents (role division, orchestrator, handoffs between agents)—and production safety (least privilege, human-in-the-loop, monitoring, evaluation).
Work too complex for one agent is handled by multi-agent, splitting roles and collaborating. For example, "research," "drafting," and "review" agents, with an orchestrator assigning tasks and integrating results. Design handoffs where agents pass work between each other; letting each focus on a small task it is good at improves quality, maintainability, and testability.
3.2.1Orchestration patterns
- Orchestrator-led: a central orchestrator dispatches to specialized agents and integrates results.
- Sequential (handoff): a pipeline passing work A → B → C in order.
- Parallel: run independent subtasks across agents concurrently, then merge.
3.2.2Production safety
Agents—especially ones that autonomously execute tools—require safety by design. (1) Least privilege: minimize each agent’s tools and access (consistent with Chapter 1 RBAC). (2) Human-in-the-loop: require human approval for high-impact actions like payments/deletions. (3) Monitoring and evaluation: continuously watch run logs, cost, success rate, and safety, and use evaluation to catch quality regressions. (4) Block harmful in/out with Content Safety and defend against prompt injection. These operationally implement Responsible AI (reliability & safety, accountability).
Scenario: a support multi-agent. A classifier judges urgency/type → a researcher searches internal knowledge via RAG → a responder drafts an answer → a reviewer checks policy compliance. The orchestrator controls the flow; high-value actions like refunds go through human-in-the-loop, each agent has least privilege, and all runs are recorded in monitoring logs.
Watch out: (1) multi-agent ≠ always optimal—simple tasks are fine with a single agent or generative AI; use it when complexity warrants. (2) Distinguish the orchestrator (control) from specialized agents (work). (3) Safety (least privilege, human-in-the-loop, monitoring, Content Safety) is a prerequisite, not optional. (4) Going multi-agent does not remove the need for permission management and monitoring—if anything, more so.
Common: (1) "split complex work across specialized agents and collaborate" = multi-agent; coordinator = orchestrator; passing work = handoff. (2) human approval before high-value/irreversible actions = human-in-the-loop. (3) per-agent permissions = least privilege (RBAC). (4) catch quality regressions = evaluation; harmful output = Content Safety. (5) multi-agent is overkill for simple tasks.
3.2.3Section summary
- Multi-agent = split complex work across specialized agents; an orchestrator coordinates and handoffs connect them
- Patterns: orchestrator-led / sequential (handoff) / parallel
- Safety: least privilege / human-in-the-loop / monitoring & evaluation / Content Safety / prompt-injection defenses
- Multi-agent is overkill for simple tasks; adopt when complexity warrants
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. What is the configuration where complex work is split across specialized collaborating agents?
Q2. In a multi-agent setup, which role dispatches tasks to specialized agents and integrates results?
Q3. Which safety practice has a human approve high-impact actions before an agent auto-executes them?
Q4. Which statement about adopting multi-agent is most appropriate?
Q5. Which is the most appropriate policy for granting an agent’s tools and access?
Q6. Which orchestration pattern passes work in order A → B → C between agents?