Interactions are diverse
The relationship between humans and agents ranges from quick, transactional requests to complex, long-term handoffs — and a robust system must support all of them. The five patterns sort naturally by who initiates the interaction: two are human-initiated, three are agent-initiated.
Key concept · Building trust is the ultimate goal
Every pattern, in a different way, is designed to build and maintain user trust — by ensuring the agent understands intent, acts in alignment with goals, provides transparency when needed, and operates securely on the user's behalf. The aim is to move from creating mere tools to designing true AI collaborators that slot safely into human workflows.
When the person makes the first move
Two patterns cover the human starting the interaction — differing in scope and duration.
Human Calls Agent
A quick, transactional request — the user needs a specific piece of information or a single action, and the agent returns it. Short-lived and stateless in spirit.
Human Delegates to Agent
A complex, long-running handoff — the user hands over a multi-step goal and the agent works autonomously toward it over time, reporting back. The interaction is a delegation, not a single call.
When the agent reaches out
Three patterns cover the agent starting an interaction — with a human, a peer agent, or an external system.
| Pattern | What it does |
|---|---|
| Agent Calls Human (HITL Escalation) | The agent recognizes its own limits or an ambiguous/high-stakes situation and escalates to a human — safety and judgment, by design. |
| Agent Delegates to Agent | A primary agent quietly delegates sub-tasks to specialists; the multi-agent complexity stays hidden so the workflow feels simple and coherent to the user. |
| Agent Calls Proxy Agent | The agent interacts with an external system in a different security context through a proxy that mediates the boundary safely. |
Key concept · Seamless collaboration hides the machinery
Users should never be burdened with the internal complexity of a multi-agent system. Agent Delegates to Agent and Agent Calls Proxy Agent are the blueprints for sophisticated, multi-part workflows that remain simple and coherent from the user's point of view — the orchestration happens behind the curtain.
Chapter 8 quiz
Fifteen questions on the interaction spectrum, the human- and agent-initiated patterns, and the role of trust. Answer first, then expand Show answer.
Part A · Multiple choice
How does the chapter frame the human-agent relationship?
- A As a single, uniform interaction
- B As diverse — from quick transactional calls to long-running delegations
- C As always human-initiated
- D As always fully automated
Show answer
B. The relationship is not monolithic; a robust system supports multiple modes.
The "Agent Calls Human" pattern should be understood as:
- A A system failure to avoid
- B A core safety feature — the agent knowing its limits and escalating
- C A performance bug
- D A security breach
Show answer
B. Escalation is a core feature, not a failure; human-in-the-loop is part of the architecture.
Which pattern fits a user needing a single, specific piece of information right now?
- A Human Delegates to Agent
- B Human Calls Agent
- C Agent Calls Proxy Agent
- D Agent Delegates to Agent
Show answer
B — Human Calls Agent. A quick, transactional interaction, versus a long-running delegation.
Which pattern handles an agent interacting with an external system in a different security context?
- A Agent Calls Proxy Agent
- B Human Calls Agent
- C Agent Calls Human
- D Single Agent Baseline
Show answer
A — Agent Calls Proxy Agent. A proxy mediates the security boundary.
What is the purpose of "Agent Delegates to Agent" from the user's perspective?
- A To expose all internal complexity to the user
- B To keep a multi-part workflow simple and coherent while specialists handle sub-tasks
- C To slow down responses
- D To remove the human entirely
Show answer
B. The orchestration is hidden so the experience stays seamless.
"Human Delegates to Agent" is best characterized as:
- A A one-off question
- B A complex, long-running handoff of a multi-step goal
- C An escalation to a human
- D A security proxy call
Show answer
B. The agent works autonomously toward the goal over time.
Part B · True or false
A well-designed system should force users to manage the internal complexity of its agents.
Show answer
False. Collaboration should be seamless; complexity stays hidden from the user.
Keeping a human in the loop is a fundamental mechanism for safety and handling ambiguity.
Show answer
True. The Agent Calls Human pattern makes HITL a core part of the architecture.
All five interaction patterns ultimately serve to build and maintain user trust.
Show answer
True. Understanding intent, aligning with goals, transparency, and secure action all support trust.
Every human-agent interaction is a long, complex delegation.
Show answer
False. Many are quick transactional calls (Human Calls Agent); the system must support both extremes.
Part C · Short answer
Name the two human-initiated interaction patterns.
Show answer
Human Calls Agent (transactional) and Human Delegates to Agent (long-running).
Name the three agent-initiated interaction patterns.
Show answer
Agent Calls Human (HITL Escalation), Agent Delegates to Agent, and Agent Calls Proxy Agent.
What single goal do all the interaction patterns share?
Show answer
Building and maintaining user trust — turning tools into safe, effective AI collaborators.
In HITL escalation, what two situations most warrant handing off to a human?
Show answer
When the agent reaches the limits of its knowledge/authority or faces an ambiguous or high-stakes decision.
What distinguishes "Human Calls Agent" from "Human Delegates to Agent"?
Show answer
Scope and duration — a Call is a quick, specific transactional request; a Delegation is a complex, long-running, multi-step goal handed over to the agent.