AssayAssay
01 / 09

Lesson 01 · Assess and bound · 13 min

From demo to deployment

Every course before this one ends at the same place: an agent that completes its task. This course starts there, because completing the task was never the hard part of running one. An agent in production reads data you did not write, acts on systems other people depend on, and does both at machine speed while nobody is watching. The discipline of managing that is agent operations.

Start with what can actually go wrong. Not vaguely — the framework this course follows names five concrete harm types, and every control in the next eight lessons exists to stop one of them.

Erroneous actionsThe agent does the wrong thing: books the wrong date, resets the wrong account, ships flawed code. The cost depends entirely on the domain.
Unauthorised actionsThe agent does something outside its permitted scope — an action that policy said needed a human, taken without one.
Biased actionsModel bias becomes real-world unfairness once actions touch people: screening, procurement, disbursement.
Data breachesAgents hold broader access than a chat window ever did, so a leak is no longer just an unfortunate reply — it can be an export.
Disruption to connected systemsA malfunctioning agent is a traffic source: it can delete, flood, or corrupt the systems it touches.

Which of these you should worry about, and how much, is not a feeling. Risk is likelihood times impact, and the framework gives each side concrete factors. Impact rises with the domain's error tolerance, access to sensitive data, access to external systems, the breadth of actions available, and — the one engineers most often skip — reversibility. Likelihood rises with autonomy, task complexity, exposure to systems you do not control, third-party opacity, and overall system complexity.

The four dimensions of governing an agent1 · Assess and boundrisk, tiers, limits, identity2 · Humans accountableowners, checkpoints, audits3 · Technical controlstest, monitor, manage change4 · End-user responsibilitytransparency, trainingre-assessdeploy only what survives all four
IMDA's framework, in one picture. The arrows matter: an anomaly found while monitoring sends you back to re-assess, so this is a cycle, not a checklist.

The four dimensions in the diagram are the shape of this course. Assess and bound before anything runs; make specific humans accountable; put technical controls around development, testing and production; and equip the people who use the agent. The arrows are the part organisations forget — monitoring findings flow back into re-assessment, so this is a loop you run, not a form you file.

Scope

Educational, not legal or compliance advice. This course teaches an engineering discipline drawn from a governance framework; where it touches regulated decisions, the binding rules are cited so you can check them, and they vary by jurisdiction.

This course follows IMDA's Model AI Governance Framework for Agentic AI — a free public document collecting practices from real deployments at banks, pharma companies, government agencies and software vendors. It is guidance rather than law, but its four dimensions map onto binding regimes: automated decisions that significantly affect people carry legal duties in the EU, and several of the use cases here are named 'high-risk' under the EU AI Act. The lessons compress the framework and make every idea runnable; the original is worth an afternoon.

Sources (3) — primary references, with jurisdiction and verification dates
  1. Model AI Governance Framework for Agentic AI, v1.5 — Infocomm Media Development Authority (IMDA), Singapore

    The four-dimension framework this course follows — assess and bound, human accountability, technical controls, end-user responsibility — assembled from real deployments. Guidance, not law, but the practices map onto the binding regimes cited elsewhere.

    Singapore (guidance; widely applicable) · published 2026-05-20 · last verified 2026-08-28

  2. Regulation (EU) 2016/679 (GDPR), Article 22 — automated individual decision-making — European Union

    Gives individuals the right not to be subject to a decision based solely on automated processing that significantly affects them (e.g. automatic refusal of credit, e-recruitment), and a right to human intervention. This is the legal backbone under a required human checkpoint.

    European Union / EEA · published 2016-04-27 · last verified 2026-08-28

  3. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III — European Union

    Classifies as 'high-risk' the AI uses this course touches: creditworthiness scoring, risk assessment and pricing for life and health insurance, and CV screening / recruitment. High-risk obligations (logging, human oversight, accuracy) become enforceable 2026-08-02.

    European Union / EEA · published 2024-08-01 · last verified 2026-08-28

In the lab, you apply the risk rubric before a model does. Six use cases, three questions each — severity, reversibility, oversight feasibility. The point of going first is that rubrics feel objective until two readers apply one.

Checkpoint

Take one agent idea from your own work and score it against the five impact factors. Which single factor moves its risk the most?

The rubric — you and the model both apply this

Assign a tier to each use case using three questions from the framework: severity — if the agent gets this wrong, how bad is it?; reversibility — can the action be undone cheaply?; oversight feasibility — is it realistic for a human to review each instance? Tier 1: low severity and fully reversible — the agent acts alone and a human audits samples afterwards. Tier 2: moderate severity or partially reversible — the agent prepares the action but a human approves before it runs. Tier 3: high severity or effectively irreversible — the agent does not perform this action at all.

Assign a tier to all six use cases using the rubric, then run the comparison. Disagreements — yours with the reference, the model's with both — are where the learning is.

How much autonomy should the agent get? Decide for all 6 before running — going first is the point.

Password resets

The helpdesk agent issues temporary passwords for verified employees who are locked out.

reversible in minuteshundreds per weekwrong target = one confused user

Firewall changes

The agent drafts and applies firewall rule changes requested by internal teams.

partially reversiblea bad rule can expose systemsa handful per day

Account deletion

The agent deletes departed employees' accounts and their data as part of offboarding.

irreversibledata loss if wronglegal retention rules apply

Goodwill refunds up to $20

A support agent issues small goodwill refunds to unhappy customers without asking anyone.

capped at $20money leaves the companyrecoverable only by asking nicely

Meeting summaries

The agent summarises internal meeting recordings and posts the notes to the team channel.

read-only inputserrors are visible and correctableposts where the attendees can see

CV screening rejections

A recruitment agent auto-rejects candidates whose CVs score below a threshold, without review.

affects people's livelihoodsbias is a known failure moderejected candidates rarely appeal