Module 3 · 55 min · 2 chapters
Finding the Work, Autonomy & Oversight
Two chapters. First decomposing real workflows into tasks, decisions, actions, handoffs and exceptions, then judging structure, reversibility, consequence and value to find bounded candidates. Then designing oversight that holds: approval gates set by action risk, permission scopes, escalation and abstention, stop conditions and kill paths, and how to avoid rubber-stamp review.
Case-led opening · fictional Harbour General
Find the work before you choose the agent.
At Harbour General, a discharge coordinator watches a shared list for patients expected home that day. She checks the EHR, looks for transport confirmation, messages community services, updates a spreadsheet and sets reminders to check again. The work is not one decision; it is a chain of small checks and handoffs that stretches across systems and time.
An agent might help carry that chain forward. But the useful question is not 'can an agent do discharge coordination?' It is 'which parts are repeatable, observable and safe to delegate, and where must a person remain responsible?'
Fictional educational case. Harbour General Hospital, its people, systems and workflow are invented for teaching purposes. This is not clinical, legal or procurement advice, and no performance outcome is claimed.
What good could look like
Capacity relief
Potentially return skilled attention from checking and chasing to exceptions and conversations that need judgement.
Lower coordination latency
Potentially start a defined follow-up when a trigger occurs rather than waiting for the next manual list review.
Fewer handoff gaps
Potentially keep a visible task state across systems, with explicit ownership and due dates.
More consistent routine work
Potentially apply the same approved sequence and verification checks, without claiming a clinical outcome.
The opportunity boundary
Coordinate
Check, request, update, verify and route bounded operational work.
Do not delegate
Interpret symptoms, determine urgency, alter medication or make a clinical decision.
Sources & evidence · 6 sources
This module cites public or consensus guidance, vendor documentation.
Content reviewed: September 2026. Publication dates of the individual sources are shown in each citation.
Anthropic, Building effective agents (19 December 2024).
A vendor engineering perspective, not independent evidence or a neutral standard; useful for the argument to start with the simplest workflow or agent architecture that meets the need.
Open sourceOpenAI, A practical guide to building agents (2025).
A vendor guide covering tools, orchestration, guardrails and bounded action design; read as an engineering perspective rather than a neutral standard.
Open sourceNIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative AI Profile.
General risk-management framing for identifying, measuring and managing risks; it does not validate any particular healthcare workflow.
Open sourceNIST AI 100-1, AI Risk Management Framework 1.0.
The Govern/Map/Measure/Manage structure behind risk-based control design, including accountability and oversight roles. It is a voluntary framework and does not prescribe specific approval thresholds.
Open sourceOWASP. GenAI Security Project — excessive agency and human oversight guidance (moving community resource).
Describes excessive agency as a distinct risk of tool-using systems and discusses human-in-the-loop controls for consequential actions. Community guidance, not a standard; continuously updated rather than versioned; checked 10 September 2026.
Open sourceWHO. Ethics and governance of artificial intelligence for health: guidance on large multi-modal models (2024).
International guidance on human oversight, accountability and appropriate use of AI in health systems. It sets principles and expectations rather than operational approval rules.
Open source