Skip to content
NTPU AI Learn

AI 302 Apply

Design the harness that keeps a real agent workflow reliable, with a Canvas

What the Canvas is for

The Agent Harness & Reliability Canvas is the capstone of AI 302. For one real use case you state the goal and the trigger, define the state model and the data each state holds, design the bounded execution loop, place checkpoints and plan how a run resumes, add verification gates and human control points, classify likely failures and set retry and recovery rules, keep the context small with compaction, decide what to observe and which reliability metrics to track, and set the operational boundaries. Human, AI, Human: a person defines the task, the agent assists or executes inside the harness, and a person reviews and stays responsible.

What you fill in

  • Workflow / System Goal and Runtime Trigger
  • State Model and State Data
  • Execution Loop
  • Checkpoints
  • Verification Gates and Human-in-the-Loop
  • Failure Taxonomy, Retry Policy and Recovery & Fallback
  • Context Compaction
  • Observability and Reliability Metrics
  • Operational Boundaries, Final Output, and the reflection question

Warning

A high-impact or irreversible action should never run on the harness alone. State which actions a person must approve before they run, and which actions the agent must never take.

Examples

The canvas page shows a worked example (a weekly supplier-invoice run) and ideas for other occupations: customer support, literature scanning, HR onboarding, financial reporting and administrative operations. All use simulated data.

Before you submit

Warning

Do not enter secrets, credentials, personal data, client data or sensitive information you are not authorized to share. Use simulated or public examples only.

What comes next

AI 302 teaches how an agent system runs reliably: it keeps state, pauses and resumes, verifies important actions, recovers from failure, and stays observable and controllable. The curriculum continues with industry project courses that apply these skills to a real organizational problem. See the course catalog for what is currently available.

Academic mapping: M5276 Week 14

AI 302 maps primarily to Week 14 of M5276 Agentic AI Innovative Applications: Agent Harness and Loop Engineering for Durable Workflow Automation: LangGraph, State Machines, Checkpointing, Context Compaction, Verification, Human-in-the-Loop, and Failure Recovery. LangGraph is used only as one implementation example, and the concepts stay framework-independent.

Planning and reasoning belong to AI 202, and role design and orchestration belong to AI 301, so this course does not repeat them.

Go to the canvas

What you need to complete

  • Complete all eight AI 302 modules
  • Score at least 80% on the course post-test
  • To earn the Agent Harness Engineering Capability Certificate, also submit an Agent Harness & Reliability Canvas and have it reviewed and passed by a Credential Admin against the rubric

How you are assessed

A Credential Admin scores the Agent Harness & Reliability Canvas against six weighted domains: Goal, Trigger & State Design; Execution Loop & Checkpointing; Verification & Human-in-the-Loop; Failure, Retry & Recovery Design; Context Compaction; and Observability, Reliability & Boundaries. A score of at least 15 out of 18, with no critical-fail rule triggered, is a PASS; 12 or below, or any triggered critical-fail rule, is a FAIL.

Info

This scoring logic is published so you know what is being assessed. It is not the same as revealing quiz answers, and it is a separate mechanism from the automatic advisory score (scoreHrCanvas) shown right after you submit.

Practical evidence

  • A recorded rubric result (PASS, REVISION_REQUIRED, or FAIL) with no critical-fail rule triggered
  • A PASS records capability evidence separately for each of the eight fine-grained AI 302 competencies (Harness vs. Agent Logic, State & Loop Design, Checkpointing & Resume, Verification & Human-in-the-Loop, Failure Classification & Recovery, Context Compaction, Observability & Reliability, and Canvas Design)

What you can earn

AI 302 Course Completion Certificate: earned once you complete all eight modules and meet the post-test threshold, with no Agent Harness & Reliability Canvas submission required. Agent Harness Engineering Capability Certificate: requires a separately submitted and reviewed Canvas. The two are independent and do not affect each other. (This certificate is a separate, parallel mechanism from AI 302's existing legacy certificate, which requires all seven practice activities and at least one canvas; neither replaces the other.)

What this credential does not mean

This credential confirms the holder completed AI 302's required course and post-test, or (for the certificate) demonstrated state design, checkpointing, verification, failure recovery, context compaction, and observability on the Agent Harness & Reliability Canvas.

  • It does not mean the holder holds a formal systems-reliability or DevOps professional license or is qualified for a specific role
  • It does not mean any specific agent harness the holder designs later is automatically correct or safe
  • It does not mean human review can be skipped before a high-impact decision (Human, AI, Human)

What next?