AIT-AGT-01
Agent workflow design
Defines an agent's objective, state, tools, stopping conditions, and orchestration pattern.
Loading your workspace…
Loading your workspace…
AI & TECHNOLOGY · COMPLETE DRAFT · READY FOR REVIEW
A technical course for designing and evaluating agents that plan, call tools, maintain state, and pause for accountable human approval.
GOVERNANCE GATE
This version exposes the proposed sources, competencies, instruction, evidence requirements, and rubric for accountable review. It cannot be enrolled in, completed, or used to issue a credential until an authorized reviewer approves a final immutable version and the program is separately published.
Status: Complete draft awaiting accountable instructional review
Draft version: 1.0.0-draft.1
Estimated learning time: 210 minutes
Program level: Course
Evidence activities: 3
Credential pathway: Proposed Certificate of Completion after review, publication, learner evidence approval, and separate administrator issuance.
LEARN → APPLY → DEMONSTRATE → REVIEW → VERIFY
AIT-AGT-01
Defines an agent's objective, state, tools, stopping conditions, and orchestration pattern.
AIT-AGT-02
Designs narrow tool contracts, permissions, validation, idempotency, and human approval for consequential actions.
AIT-AGT-03
Evaluates trajectories, tool choices, outcomes, recovery, and operational risk using preserved traces.
CONTINUING PROFESSIONAL CASE
A company wants an agent to investigate account issues, search internal policy, draft explanations, and request refunds through a tool. A prototype sometimes chooses the wrong account and repeats a refund call after a timeout.
Constraint: The first release must assist staff, preserve a complete trace, and prevent duplicate or unauthorized financial actions.
Learner task: Design the agent loop, tools, permissions, human approvals, recovery behavior, and trajectory evaluation plan.
FIVE SOURCE-GROUNDED LESSONS
LESSON 1
Purpose: Define bounded autonomy as an explicit control loop.
An agent observes state, selects an action, receives a result, updates state, and continues until a stopping condition. Goals and limits must be operational, not merely aspirational.
Professional example: The support agent stops when identity is uncertain, policy conflicts, or a refund requires approval.
Evidence connection: Current Agents SDK guidance defines agents around instructions, models, tools, state, orchestration, and results.
Diagram the support agent's loop, terminal states, and conditions that force human handoff.
What most clearly defines an agent loop?
Expected: Repeated state-aware action selection with explicit stopping conditions. — Agent behavior arises from the loop, state, actions, and termination rules.
LESSON 2
Purpose: Make actions narrow, validated, and reviewable.
A tool should expose the smallest useful capability, validate structured arguments, authorize the actor, return explicit results, and handle retries without duplicating consequential effects.
Professional example: A refund request tool creates a pending proposal keyed by request ID rather than issuing money directly.
Evidence connection: Agent tooling documentation treats function schemas as contracts; risk guidance requires controls proportional to effects.
Redesign the refund tool with narrow scope, stable IDs, authorization checks, and a pending approval state.
Which tool design is safest?
Expected: proposeRefund(requestId, accountId, amount, reason) validates and awaits authorized approval — Narrow contracts and staged authority reduce unintended effects.
LESSON 3
Purpose: Preserve relevant context without confusing ownership.
Agent state should record facts, decisions, tool results, and pending approvals needed for continuation. Handoffs need clear ownership and should not silently expand permissions.
Professional example: A billing specialist receives verified account context and the unresolved policy question, but not an unrestricted tool set.
Evidence connection: Agents SDK documentation separates result state, resumability, specialist definitions, and orchestration choices.
Define the state schema and handoff contract between triage and billing specialists.
What belongs in durable agent state?
Expected: Verified facts, tool results, decisions, ownership, and pending approvals needed to resume safely. — Useful state is scoped, attributable, and sufficient for safe continuation.
LESSON 4
Purpose: Block or pause unsafe actions at the right boundary.
Guardrails can validate inputs, outputs, and tool use, but consequential actions require explicit authorization, affected-user protections, and fail-closed behavior when evidence is missing.
Professional example: The agent may draft a refund explanation but cannot submit the refund until staff review the account, amount, and policy basis.
Evidence connection: Agent safety guidance emphasizes guardrails and human review; NIST frames human oversight and incident response as governance controls.
Place guardrails and approval pauses around identity, policy, and refund execution.
Where should human approval occur?
Expected: After the exact proposed action is visible and before the financial effect. — Approval must be informed, specific, and timely.
LESSON 5
Purpose: Evaluate how the agent acts, not only its final answer.
A correct final response can conceal unsafe tools, needless loops, or accidental effects. Evaluation should inspect tool selection, arguments, state changes, approval compliance, outcome, cost, and recovery from failure.
Professional example: A timeout test verifies that retrying the proposal tool does not create a second refund.
Evidence connection: Current agent evaluation guidance focuses on workflow-level behavior and traces rather than final text alone.
Build success, ambiguity, timeout, denial, and duplicate-action scenarios for the support agent.
Which evaluation is most complete?
Expected: Inspect the trajectory, tool arguments, effects, approvals, outcome, and recovery. — Agent quality includes the path and its effects.
ASSESSED APPLICATION
PROPOSED REVIEW RUBRIC
Problem framing and operational boundary is explicit, workable, and supported by relevant evidence.
Competency: AIT-AGT-01
Technical design and implementation rationale is explicit, workable, and supported by relevant evidence.
Competency: AIT-AGT-02
Evaluation design and preserved evidence is explicit, workable, and supported by relevant evidence.
Competency: AIT-AGT-03
Risk controls and accountable governance is explicit, workable, and supported by relevant evidence.
Competency: AIT-AGT-03
Failure analysis, revision, and professional judgment is explicit, workable, and supported by relevant evidence.
Competency: AIT-AGT-03
PRIMARY SOURCE BASE
OpenAI · OpenAI API Documentation · Current documentation accessed August 2026
Provides current primary documentation for agent definitions, tools, state, orchestration, guardrails, tracing, and evaluation.
OpenAI · OpenAI API Documentation · Current documentation accessed August 2026
Provides current primary guidance for defining objectives, datasets, metrics, and continuous evaluation of AI applications.
National Institute of Standards and Technology · NIST AI 600-1 · 2024
Provides a cross-sector framework for identifying, measuring, and managing generative-AI risks across the lifecycle.