Architecture blueprint
Design the agent loop: goal → plan → tool selection → action → observation → validation → next step or escalation. Keep the permissions and tool scope narrower than the business goal wherever possible.
Pathway 3 · Lesson 0590 min
AI Agent Architecture
Designing autonomous AI systems that take action
Learning Objective
By the end of this lesson, you will be able to: Design an AI agent architecture with appropriate tools, memory, and oversight structures, and identify the failure modes that require human intervention.
82%
of executives plan to adopt AI agents within 3 years
Engage
"AI agents don't just answer questions — they take action. They browse the web, write code, send emails, and make API calls. This is transformative. It's also dangerous if not designed carefully."
Explore
What AI agents are, how they work, and why they represent a fundamental shift in how AI is used — from generation to action.
Explain
What Makes an AI Agent
An AI agent is a system that can:
The key difference from a standard LLM: agents take action in the world, not just generate text. This makes them dramatically more powerful — and dramatically more risky if not governed correctly.
Agent Architecture Patterns
| Pattern | How It Works | Best For | Risk Level |
|---|---|---|---|
| ReAct | Reason → Act → Observe → Repeat | Single-agent, well-defined tasks | Medium |
| Plan-and-Execute | Plan all steps first, then execute | Complex multi-step tasks | Medium-High |
| Multi-Agent (Supervisor) | One agent orchestrates specialist agents | Parallel complex tasks | High |
| Multi-Agent (Peer) | Agents collaborate as equals | Creative and research tasks | High |
| Human-in-the-Loop | Human approves before each action | High-stakes, irreversible actions | Low |
⚠️
Agent Failure Modes to Design Against
1. Hallucinated actions: Agent invents tool calls that don't exist 2. Infinite loops: Agent gets stuck repeating the same action 3. Scope creep: Agent takes actions outside its intended domain 4. Irreversible mistakes: Agent deletes data, sends emails, or makes purchases that can't be undone 5. Prompt injection: Malicious content in the environment hijacks the agent's instructions Every agent design must explicitly address how to detect and recover from each of these.
Explain
Governance: Human Oversight Structures
Three oversight models for AI agents:
Human-in-the-loop: Human approves every action before execution. Maximum safety, minimum autonomy. Use for: high-stakes, irreversible, or novel actions.
Human-on-the-loop: Agent acts autonomously; human monitors and can intervene. Use for: well-defined, reversible tasks with clear success criteria.
Human-out-of-the-loop: Fully autonomous. Use only for: extremely well-tested, low-stakes, fully reversible tasks with comprehensive monitoring.
The principle: Start with human-in-the-loop. Reduce oversight only as trust is established through demonstrated reliability.
Try It · Reflect & Apply
Design an agent: Choose a multi-step task in your organisation that could benefit from an AI agent. Define: What tools would it need? What oversight model would you use? What are the top 3 failure modes to design against? What would trigger human escalation?
Your notes are saved in your browser for this session.
Evaluate · Knowledge Check
1 / 3What fundamentally distinguishes an AI agent from a standard LLM?
Builder Engineering Studio
Agent Reliability Studio
Engineering trade-off
Autonomy reduces handling time but increases the blast radius of mistakes. Start with human approval for irreversible actions, then earn autonomy through evidence of reliable performance.
Quality gate
Define hard stops for money movement, external communication, sensitive-data exposure, destructive operations, and low-confidence outputs. Simulate these before production access is granted.
Evidence to measure
Measure task success, tool-call error rate, recovery rate, human intervention rate, loop frequency, policy violations, and time-to-escalation for abnormal behaviour.
Build task
Create an agent permission matrix: permitted tools, maximum action scope, approval thresholds, audit records, rollback method, and accountable human owner.
External video references
Keep Learning on YouTube
Open in a new tab
AI Agents ExplainedEmbedded lesson videoWhat AI agents are, how they work, and why they represent a fundamental shift in how AI is used — from generation to action.Watch · Watch on YouTube Complete Agentic AI CourseFree technical courseA long-form implementation companion covering agents, tools, RAG, orchestration, and production decision points.Watch · Extended course
Human-centred work design
Human Oversight Check
Use this original practice tool to test whether an AI-enabled workflow supports real human judgment, rather than merely placing a person at the end of an automated process.
Self-check
0/5
Start by making the work and decision rights explicit
Calibrate autonomy
Match AI freedom to consequence. Low-consequence drafting may run quickly; high-consequence decisions need explicit human approval.
Keep context visible
Make the objective, inputs, constraints, source material, and current system state easy to inspect before anyone approves an output.
Design active review
Review should require informed judgment, a plausible override, and the authority to act—not a symbolic sign-off after the fact.
Route exceptions clearly
Specify what uncertainty, conflict, missing evidence, or risk looks like—and who receives the escalation with enough context to decide.
Maintain human capability
Protect deliberate practice in the work that people may one day need to verify, challenge, or perform without AI assistance.
Apply it to one workflow
Ready to move on?
Mark this lesson complete and continue to "Architecting the Hybrid Workforce".