What is an agent? About 13 min · includes 7 min video

A workflow is the connected sequence of steps used to complete the work. An agent may decide which permitted step is appropriate next. Conventional automation follows a sequence fixed in advance.

Many organisational systems are hybrids. Fixed rules can control predictable or consequential steps while an AI agent handles interpretation, information gathering or preparation.

What must be designed into a simple agent? About 7 min

A simple agent brings together a small number of building blocks. Each one answers a different design question.

What makes a simple agent suitable? About 9 min · includes 5 min reading

From information to action: how much authority should the agent have? About 6 min · includes 1:45 video excerpt

Agents can support work at different levels of authority. The key design question is what the agent may do independently and what remains subject to human judgement or approval.

How do design decisions become agent instructions? About 8 min

1. Which six elements must be designed before the tool is configured?

2. How is the design translated into instructions?

3. How is the design configured and refined?

When behaviour is poor, diagnose the design: is the goal unclear, information missing, process underspecified, output ambiguous or authority excessive?

What can you use to build a simple agent? About 5 min · choose one 2 min reading

No-code builders configure an existing model through instructions, knowledge and selected capabilities. They are well suited to bounded prototypes where you can inspect the behaviour without programming.

What starts and ends an agent run? About 4 min

An agent needs a defined trigger that begins the work and explicit conditions that end it. A trigger starts a run; it does not expand the agent’s authority.

When should the run stop?

How does the designed agent operate? About 3 min

How does the agent continue through the task?

What does one complete agent run look like?

Follow the recurring service enquiry from the initial report to a human-controlled outcome.

How well do you understand simple agents? About 5 min

Use these four unmarked knowledge checks to apply ideas from across the week. Select one answer for each question to receive immediate feedback.

What makes an agentic system complex? About 10 min · includes 7 min video

Where does the complexity come from?

Diagnose the work in this order. Each question narrows the design problem before you consider how many agents to use.

1. How much work must be connected?

   Look for several tasks, information sources, tools or decision points that must contribute to one outcome.

   In this scenarioTechnical history, billing information, customer impact and remediation guidance must be brought together.

2. How much can the pathway vary?

   Look for branches, changing priorities, uncertainty, retained context or different specialist knowledge depending on the case.

   In this scenarioThe appropriate response changes with the technical cause, disputed period, impact on the customer and available remedies.

3. Where do authority and responsibility change?

   Look for restricted information, separate permissions, consequential decisions, handovers or points where a person must intervene.

   In this scenarioTechnical support, billing remediation and vulnerability escalation may have different information access and decision authority.

Could the existing agent handle the additional work using a skill? About 6 min

Week 7 introduced a skill as a reusable capability that can package instructions, approved knowledge, tools and a repeatable procedure. Week 8 uses that idea as an architecture choice: before creating another agent, ask whether the required specialisation could become a skill used by the existing agent.

How could each support the same customer enquiry?

A customer reports recurring internet disruptions, questions their bill and needs a recommended next step. Tools, skills and agents can work together within the same service.

Consider: if technical and billing work uses the same information and permissions, what would you examine before introducing separate agents?

Why might a capability become a skill?

What would a skill need to specify?

---
name: recurring-service-enquiry
description: Use when a frontline employee receives a report of repeated internet disruptions.
---

1. Gather the required service details.
2. Identify important missing information.
3. Consult only the approved troubleshooting guidance.
4. Prepare a structured recommended next step.
5. Escalate when the guidance does not cover the situation.

When would a separate agent still be justified?

Consider a specialist agent when the work requires a separate goal, materially different instructions or context, distinct permissions, independent execution, a separate review point or evidence that specialisation improves quality, reliability or control.

Optional reading, about 2 min: read only “Risks and safety”, concentrating on approval, trust and the permissions available to a skill. This extension is excluded from the required pathway time. Open the safety section ↗

How does the case evidence change the architecture decision? About 22 min · includes 10 min reading and 5 min video

1 · Test the single-agent baseline

Mars-rover decision support

One agent received each complete synthetic scenario and produced a recommended action, risk level, hazards, reasoning and confidence rating. It did not need external tools or handovers between specialists.

Architecture implication: complex reasoning can remain with one agent when the task is coherent, the full context fits together and separate responsibilities add little value.

Required reading, about 3 min: Abstract, “Experimental setup” and the first two paragraphs of “Discussion”.

Open the focused research reading ↗

Case check 1 of 4

What is the strongest architectural lesson from the Mars-rover case?

A complex task always performs best with one agent.Complex reasoning alone does not establish a need for multiple agents.A single agent is appropriate only when no judgement is required.

2 · Increase workload diversity and scale

Scaling mixed clinical task batches

A controlled study compared one agent handling mixed batches of retrieval, clinical-note extraction and medication-dosing tasks with an orchestrator assigning tasks to specialist workers. As batch size increased from 5 to 80 tasks, reported single-agent accuracy fell from 73.1% to 16.6%, while multi-agent accuracy fell from 90.6% to 65.3%.

Architecture implication: orchestration becomes more plausible when mixed tasks compete for context and attention. These curated tasks had deterministic answers and were not a deployed clinical service.

Required reading, about 3 min: Abstract, the architecture figure and the opening of “Discussion”.

Open the focused research reading ↗

Case check 2 of 4

What can the clinical-workloads study reasonably support?

Multi-agent systems are ready to make consequential clinical decisions.Every sufficiently large task batch should be divided among specialist agents.Specialist orchestration may help with larger mixed-task batches in this controlled setting.

3 · Separate a repeatable workflow

UK utilities email response

A small pilot separated email categorisation, summary, template selection, knowledge retrieval and response generation into specialist components before employee refinement. The workflow contained recurring stages with different instructions and information needs.

Architecture implication: multiple specialists may be justified when responsibilities can be bounded and coordinated. The pilot involved only three employees and 24 emails, so its reported results do not establish performance at organisational scale.

Continue to the focused case reading ↓

FOCUSED CASE READING · APPROX. 4 MIN

UK utilities customer-service automation

Read Section 5.4 for the workflow, pilot and results, then Section 5.5 for why the authors associate the hierarchical pattern with a stable, policy-constrained pipeline.

Open the targeted sections ↗

READ WITH CAUTION

What should you extract?

• Which responsibilities were separated?
• What information moved between components?
• Where did employee refinement remain?
• How small was the reported pilot, and what does that limit?

Harri Renney, Maxim Nethercott, Nathan Renney and Peter Hayes, Journal on Artificial Intelligence, 8, 231–257. DOI ↗

Case check 3 of 4

How should the UK utilities pilot influence an architecture decision?

Use it as limited evidence that bounded specialist components can support a repeatable workflow.Treat its reported accuracy as proof that the architecture will perform reliably at organisational scale.Reject specialist architectures because employee refinement remained in the workflow.

4 · Examine an architecture demonstration

High-pressure coordination

A vendor-hosted conference session illustrates how distributed inputs, specialist contributions and human monitoring could be coordinated while preserving a shared view of a high-pressure situation.

Architecture implication: the demonstration helps reveal routing, shared-context, oversight and recovery requirements. It is an illustration of a proposed architecture rather than independent evidence of operational effectiveness.

REQUIRED FOCUSED VIEWING · APPROX. 5 MINUTES

Enhanced decision-making in disaster response with multi-agent AI

Watch 22:02–23:21 for the concurrent inputs, multi-agent integration and simulation layer. Then watch 27:41–31:35 for orchestration, specialist agents, shared outputs, human monitoring and transparency.

Watch on NVIDIA On-Demand ↗

VIEW WITH CAUTION

What should you extract?

• What does the orchestrator coordinate?
• What do the specialist agents contribute?
• What information moves between them?
• Where does a person remain responsible?

This is a vendor-hosted conference case, and the published transcript is AI-generated. Use it to examine the presented architecture and acknowledged limitations, rather than as independent evidence of effectiveness.

Case check 4 of 4

What is the most defensible use of the high-pressure coordination case?

Use it as independent proof that the demonstrated system improves disaster-response outcomes.Use it to identify coordination requirements that would need operational validation.Use it to conclude that every high-pressure decision should be distributed across several agents.

Which architecture is justified? About 8 min · includes 5 min reading

Architecture comparison adapted from OpenAI, A Practical Guide to Building Agents, and Renney et al. (2026).

How can multiple agents be coordinated? About 5 min

After specialist agents are justified, the architecture must specify how work is routed, what each role receives and produces, how control moves and how the overall outcome remains coherent.

Terminology: coordinator is the generic role responsible for directing or reconciling work. An orchestrator is a coordinating component. A manager pattern keeps control with a central agent, while a handover pattern transfers active control.

What must be defined at every handover?

A handover needs an explicit interface contract so that each participant knows what it will receive, what it must produce and what happens when the transfer is incomplete.

How does a coordinated system remain coherent? About 2 min

Each component may complete its own task while the overall system still fails. The architecture therefore needs a shared view of progress, bounded recovery behaviour and an inspectable record of the complete workflow.

Where should context and state live?

What coordination failures must the design handle?

What should be visible across the complete workflow?

How well do you understand complex agents? About 5 min

Use these four unmarked knowledge checks to apply ideas from across the week. Select one answer for each question to receive immediate feedback.

⚖What do Australia’s AI Ethics Principles require? About 11 min · includes 4 min reading

Responsible AI applies beyond generative AI. It concerns how organisations design, use and govern AI-enabled and automated systems, including the people, processes and decisions around them.

What shapes a responsible organisational decision?

An organisation should state the intended benefit before accepting a risk. It must then consider whether the proposed use, affected stakeholders and possible harms are proportionate to that benefit.

Who does the AI serve?

The stated user is only one stakeholder. Examine how benefit, decision power and consequences are distributed before deciding that the proposed purpose is responsible.

Read the visual clockwise: begin with wellbeing and human-centred values, then examine fairness, privacy, reliability, transparency, contestability and accountability. The principles overlap and must be interpreted together for the organisational context.

Focused reading, about 4 min: read “About the AI Ethics Principles” and the eight principle descriptions from “Human, societal and environmental wellbeing” through “Accountability”. Skip the archived development resources. Open the targeted principles ↗

→How do principles become responsible requirements? About 6 min

Principles become actionable only when an organisation turns them into a sequence of decisions. Use the six questions below as the anchor for every case in this week. Each answer should produce something concrete that can be examined, assigned and improved.

1. 1
   Purpose and principleWhat benefit is sought, and what responsible outcome must be protected?
   State the benefit before accepting the risk. Select the materially relevant principles and explain what they require for this use.Output: an intended benefit and responsible-AI objective.
2. 2
   People and powerWho benefits, who decides and who bears the consequences?
   Look beyond the stated user. Include people affected directly or indirectly, groups carrying greater risk and perspectives missing from the design.Output: a stakeholder and power map.
3. 3
   Failure and harmHow could the system produce an unacceptable outcome?
   Trace the pathway rather than naming a generic risk. Locate where bias, unsupported output, lost uncertainty, misuse or overreliance could enter, then describe who is harmed and how.Output: a specific failure-to-harm pathway.
4. 4
   ControlsWhat will prevent, detect and respond to it?
   Choose controls that interrupt the pathway to harm. A control should have a clear mechanism. Training or a policy statement alone may be insufficient.Output: preventive, detective, responsive and corrective controls.
5. 5
   OwnerWho has authority and accountability?
   Assign ownership to identifiable roles. Specify who approves the use, operates each control, reviews exceptions, can stop the system and remains answerable for outcomes.Output: named decision and control owners.
6. 6
   Evidence, monitoring and redressWhat evidence will trigger action, and how will harm be remedied?
   Decide what counts as evidence before deployment. Combine system measures with human review and affected-stakeholder feedback, define thresholds, investigate failures and provide remedy.Output: measures, review triggers and a redress pathway.

Where can model behaviour become organisational harm?

Bias and confabulation require different tests. Bias is a systematic pattern that may disadvantage particular people or groups. Confabulation is a plausible output that is unsupported or incorrect. Either can become more harmful when people or automated processes rely on the output without appropriate evidence.

How does an abstract principle become something operational?

⊘How do controls interrupt the pathway to harm? About 3 min

A control is a specific mechanism that changes what the system, its users or the organisation can do. Controls should be proportionate to the harm and assigned to people with the authority and resources to operate them.

How do controls operate before, during and after harm?

Who does what?

◎How will the organisation know when action is required? About 11 min

Monitoring tests whether the intended benefit is being achieved, whether controls are working and whether unacceptable outcomes are emerging. It must connect evidence to thresholds, decisions and remedies.

What should monitoring reveal?

Can one LLM evaluate another?

Decision rule: agreement between two models establishes model agreement. Accuracy requires comparison with authoritative evidence.

☎Who is responsible when synthetic media deceives? About 8 min · includes required video and reading

AI-generated political robocall

◖Synthetic voiceA recognisable public figure was imitated.

☎Spoofed distributionThe call appeared to come from a trusted source.

!Public consequenceVoters received deceptive election information.

The situation: An AI-generated imitation of Joe Biden’s voice encouraged recipients not to vote in the 2024 New Hampshire primary. Caller information was spoofed.

What AI was used for: Synthetic voice made the message sound as though it came from a recognisable public figure.

What happened next: US regulators investigated the communications chain and took enforcement action.

Trace responsibility through the case

1 · Purpose and principleThe operational objective was to influence voter behaviour; responsible use requires transparency, informed participation and accountable senders and carriers.

2 · People and powerVoters may act on false voting instructions; candidates may be misrepresented; election officials and communications providers must address misleading calls.

3 · Failure and harmDeception may interfere with informed political participation and public trust.

4 · ControlsProposed safeguards: authenticate callers to detect spoofing, disclose synthetic speech, retain records linking calls to senders, and stop deceptive calls when detected.

5 · OwnerResponsibility must be assigned across the sender and organisations transmitting the call.

6 · Evidence, monitoring and redressDetect spoofing and complaints, preserve evidence, stop dissemination and communicate corrections.

Read, about 3 min: the opening summary, “Background” and “Enforcement Bureau Action”. Open the focused FCC reading ↗

◎Can a legitimate purpose excuse weak governance? About 9 min · includes required video and reading

Bunnings facial recognition

▣CaptureA camera image creates biometric information.

◎CompareThe system checks the image against people of security concern.

⚑Handle a matchNotice, proportionality, access and review shape responsible use.

The situation: Bunnings used facial-recognition technology for a limited security purpose.

What AI was used for: The system compared facial images to help identify people of security concern.

What happened next: The 2026 Tribunal decision accepted the limited purpose while affirming failures in notification and privacy governance.

Complete the second half of the pathway

The purpose and principle, people and power, and failure and harm are provided. Before opening the final three stages, predict what would turn these concerns into organisational action.

1 · Purpose and principleThe security purpose must be weighed against privacy, transparency, fairness and accountability requirements.

2 · People and powerCustomers face biometric collection; employees seek protection from security threats; people flagged by the system may face action based on a match.

3 · Failure and harmPeople may be identified or tracked without adequate notice or governance.

4 · ControlsWhat mechanisms would address necessity, notice, data use and review?

Reveal one response

Assess necessity and proportionality, give clear notice, restrict data use and require meaningful review.

5 · OwnerWhich organisational role should maintain these privacy and security controls?

Reveal one response

Assign identifiable privacy and security owners with authority to review and change the use.

6 · Evidence, monitoring and redressWhat evidence and remedies would reveal whether the use remains acceptable?

Reveal one response

Audit use and false matches, review complaints and provide correction or deletion pathways where required.

Read, about 3 min: the complete short OAIC statement. Open the OAIC statement ↗

▣Who owns incorrect AI guidance? About 5 min · includes required video

Air Canada chatbot

?AskA customer sought fare guidance.

▣AdviseThe chatbot supplied incorrect policy information.

!RelyThe customer acted on the guidance.

The situation: A customer sought information about a bereavement fare.

What AI was used for: A chatbot provided customer-service guidance through Air Canada’s website.

What happened next: The customer relied on incorrect guidance, and the Civil Resolution Tribunal held the organisation responsible for the information provided.

Locate organisational responsibility

1 · Purpose and principleThe service sought scalable customer guidance; reliability, transparency and accountability constrain how that benefit is delivered.

2 · People and powerCustomers, frontline staff and policy owners depend on consistent information.

3 · Failure and harmIncorrect guidance may cause financial loss and deny expected support.

4 · ControlsUse approved policy sources, conflict checks, review and clear escalation.

5 · OwnerWho must keep the guidance accurate and resolve conflicts?

Reveal one response

The organisation needs identifiable policy, service and system owners. Responsibility cannot be transferred to the chatbot.

6 · Evidence, monitoring and redressWhat evidence and remedy would reveal or correct harmful guidance?

Reveal one response

Monitor conflicting answers and complaints, correct the information, preserve the customer’s challenge pathway and provide an appropriate remedy.

Read, about 3 min: paragraphs 23–30 of Moffatt v Air Canada, 2024 BCCRT 149. Open the tribunal decision on CanLII ↗

⌾How can uncertain AI evidence become harmful? About 4 min · includes required video

Robert Williams

▧CaptureA poor-quality surveillance image was used.

⌾SuggestFacial recognition produced a possible match.

!Over-relyUncertain output was treated as stronger evidence.

The situation: Police investigated a theft using a poor-quality surveillance image.

What AI was used for: Facial recognition generated a possible match.

What happened next: Police treated the output as stronger evidence than it was, contributing to a wrongful arrest.

Preserve uncertainty through the pathway

1 · Purpose and principleThe system sought to assist identification; fairness, reliability, transparency, contestability and accountability constrain how a possible match may be used.

2 · People and powerThe person matched risks wrongful arrest; investigators risk relying on mistaken identification; the public and affected communities depend on fair, reliable investigations.

3 · Failure and harmWhere did a possible match become consequential evidence?

Reveal one response

Uncertainty was lost when a possible match was treated as identification and used to support arrest.

4 · ControlsWhat should prevent over-reliance?

Reveal one response

Require independent corroboration, communicate uncertainty and prohibit a match from serving as sole identification evidence.

5 · OwnerInvestigative leaders and system owners must govern evidence use and review.

6 · Evidence, monitoring and redressAudit false matches, review affected groups and provide correction and remedy pathways.

Key lesson: an uncertain model output must remain uncertain throughout the human decision process.

$How can automation create systemic harm? About 6 min · includes required video

Robodebt

▤InputIncome information entered an administrative process.

÷AverageAutomation inferred fortnightly income.

$Assert debtPeople faced debt-recovery action at scale.

The situation: Government debt-recovery processes affected people receiving income support.

What automation was used for: Income averaging was used within automated debt identification and recovery processes.

What happened next: The scheme contributed to unlawful debt assertions and severe impacts on affected people.

The failure involved the method, the policy authorising its use, weak attention to individual circumstances and inadequate challenge and remedy pathways.

Trace systemic responsibility

Complete the six-stage pathway before opening one defensible analysis.

1 · Purpose and principleWhich principles were placed at risk?

2 · People and powerWhose circumstances and perspectives were missing?

3 · Failure and harmHow could an administrative method create harm at scale?

4 · ControlsWhat should verify legality, evidence and individual circumstances?

5 · OwnerWho remains accountable across policy, operations and technology?

6 · Evidence, monitoring and redressWhat evidence should trigger correction, suspension and remedy?

Reveal one defensible trace

Principles and stakeholders: human-centred values, fairness, reliability, transparency, contestability and accountability were material for people receiving income support and the public institutions serving them.

Controls and ownership: legal review, evidence validation, consideration of individual circumstances and accessible challenge pathways required accountable policy, operational and technology owners.

Monitoring and redress: complaints, overturned debts and evidence of hardship should have triggered correction, suspension and remedy.

👁What makes human oversight meaningful? About 11 min · includes 5 min reading

The person needs all six conditions below. If one is missing, the organisation should question whether human review is genuinely capable of preventing or responding to harm.

1. Understand
   Know the system’s purpose, limitations and personal responsibilities.Transparency and explainability · accountability
2. See
   Receive relevant evidence, warnings, uncertainty and emerging risks.Transparency and explainability · reliability and safety
3. Decide
   Exercise independent judgement without automatically accepting the output.Human-centred values · fairness
4. Act
   Have the authority, mechanism and time to intervene, override, escalate or stop.Contestability · accountability
5. Be supported
   Receive training, manageable workloads and organisational backing.Reliability and safety · accountability
6. Be accountable
   Record and review decisions and interventions so the system can improve.Accountability · contestability

✓Can you govern an AI system responsibly? About 8 min

These questions require you to connect principles, stakeholders, harms, controls, ownership, evidence and human oversight. Choose the strongest response.