AI 302 Assessment
Show that you can design the harness that keeps an agent system reliable: state, checkpoints, verification, human control, recovery, compaction and observability
What is assessed
The assessment checks judgment, not vocabulary. It has five parts with the weights below.
| Part | Weight |
|---|---|
| Knowledge Check (12 questions on the Agent Harness, state and state machines, checkpoints and durable execution, verification and Human-in-the-Loop, retry and failure recovery, context compaction, and observability) | 15% |
| State & Checkpoint Design (choosing the harness, modelling states and transitions, placing safe checkpoints and defining resume behavior) | 20% |
| Verification & HITL Exercise (deciding what continues automatically, what is verified, what needs review or approval, and what must be observed) | 15% |
| Failure & Recovery Exercise (classifying failures, deciding retryability, choosing fallback, escalation and safe stop, and compacting context) | 20% |
| Final Application (Agent Harness & Reliability Canvas) | 30% |
Scoring and passing
Quiz scoring happens on the server. Correct answers and explanations never appear on public pages or in downloadable files. The passing threshold is 80%.
Info
Questions appear once you start the assessment.
Canvas rubric
| Criterion | Weight |
|---|---|
| Goal, trigger and state model | 15% |
| State data and execution loop | 15% |
| Checkpoints and resume behavior | 15% |
| Verification gates and Human-in-the-Loop | 15% |
| Failure taxonomy, retry policy and recovery | 15% |
| Context compaction, observability and reliability metrics | 15% |
| Operational boundaries, final output and reflection question | 10% |
Info
Reviewers use these weights when they read a Canvas. The automatic advisory score shown after you submit and in My Learning is separate: it counts each of the sixteen Canvas sections and the reflection question equally (up to 3 points each), is advisory only and does not replace human review.
- 1Excellent
The goal and trigger are clear, and states, transitions and stored data are defined with sensitive data kept out. The loop is bounded. Checkpoints say what is saved, why it is safe and what resume does, and unsafe points are named. Critical steps have a verification method, criterion, failure route and evidence, and people review or approve the high-impact points. Failures are classified, retry is limited and never applied blindly to actions with external effects, and recovery has an ordered path to safe stop. Compaction keeps decision-relevant state, observability lets a human explain a run, metrics have targets, and operational boundaries are explicit. - 2Competent
The main states, checkpoints, gates and recovery steps are designed, some are vague or leave a replay risk open, and the human controls are broadly sensible. - 3Needs improvement
State is missing or lives only in the prompt, checkpoints are absent or unsafe, loops have no limit, retries are applied to every failure, high-impact actions have no human approval, or nothing is recorded that would let someone explain a failed run.
Reflection
Write 100 to 200 words in response to this question:
Info
If this Agent fails halfway through a long-running task, can we explain what happened, resume safely, avoid duplicate actions, and keep a human in control?
Start
- Start the AI 302 assessment (post-test)
- My Learning: Progress and results