Foundation 05 · Exception design

Design the stop path
before the happy path.

A workflow becomes more trustworthy when it knows when to pause. An exception is not a model failure to hide; it is a defined handoff from bounded automation to accountable judgment.

Learning boundary. This lesson provides a design method. It does not make an autonomous system secure, compliant, reliable, or suitable for a particular environment. Human review and domain-specific controls remain necessary.

Learning objective

Convert uncertainty and boundary breaches into a deliberate, reviewable hold.

By the end of this lesson, you should be able to define exception triggers, the safe state, a named resolver, the evidence packet they need, and the maintenance change that may follow a resolved exception.

01 · Trigger

Which condition means “do not continue”?

Name scope breaches, conflicting instructions, unsafe tool responses, missing evidence, policy exceptions, unexpected outputs, and actions that change another person’s rights.

02 · Hold

What is the safe state while a decision waits?

Specify whether the workflow drafts, queues, retries, rolls back, notifies, or does nothing. A hold should avoid quietly creating a new irreversible dependency.

03 · Resolve

Who receives the exception, with what context?

Route to a named owner or role with an evidence packet: task, source identifiers, boundary, candidate actions, risk signal, and the choice that requires accountability.

04 · Learn

What changes after resolution?

Separate process improvement from automatic policy expansion. A resolved exception may require a new test, clearer documentation, tighter scope, or an explicit human decision to change the rule.

Practice lab

Design a hold for a procurement workflow.

A workflow can compare approved vendors and draft a purchase request within a low-value limit. It receives a request that matches a vendor name but has a conflicting invoice total and an unrecognized delivery location.

  1. Write the exact triggers that prevent automatic progression.
  2. Describe the safe hold state and which draft artifacts may remain visible.
  3. Name the accountable resolver and the minimum evidence packet they need.
  4. Describe how a resolved exception becomes a review item without automatically widening the workflow’s authority.

Source shelf

Model the threat and the handoff.

  1. OWASP Agentic AI — Threats and MitigationsA threat-model-based reference for emerging agentic threats and mitigations. Use it to generate questions for local review, not to claim a workflow is secure.
  2. NIST AI Risk Management FrameworkUse the framework to structure risk-management work around context, measurement, management, and governance; retain human judgment over what an exception means locally.

Complete the pathway

Return to the
resource field.