# Agents and Workflows – Track Overview
Independent preparation for the OpenAI Academy Agents and Workflows course – delegating structured work to agents in ChatGPT Work with real human oversight, across six domains and two independent mock exams.
import { Card, CardGrid, Badge } from '@prosefly/astro-components';
Agents track 6 domains 50-item mocks Foundations pathway · course 3
This track is independent preparation for the OpenAI Academy course **Agents and Workflows**, the third and final course in the **Foundations pathway** (after [AI Foundations](/openai/foundations/) and [Applied AI Foundations](/openai/applied-ai/)). It is not an official OpenAI course and not an official assessment; see the [credential landscape](/openai/credentials/) for how Academy badges differ from certifications. The material here is built from publicly available OpenAI learning objectives.
## What this track prepares you for
The Academy course **Agents and Workflows** runs **75–90 minutes**, its product focus is **ChatGPT Work**, and its published objectives are to **define tasks, context, boundaries and checkpoints, then review and improve results**. The course is about *delegating* structured work to an agent with human oversight — deciding what to hand off, briefing it well, setting the limits it must operate inside, and verifying what comes back.
This is a **non-developer track**. It teaches how to delegate work to agents inside ChatGPT Work and Codex-style surfaces as a knowledge worker, not how to build agents in code. If you want the builder's view — the Responses API, the Agents SDK and the Agents API — take the [API Developer Path](/openai/api/) instead. The two tracks share vocabulary but answer different questions: this one asks *should I delegate this, and how do I brief and check it*; the API track asks *how do I construct the runtime*.
## Blueprint
Our mock exams and domain pages follow this weighting. Item counts are approximate on a 50-item mock.
| # | Domain | Weight | Items (approx.) | Page |
| --- | --- | --- | --- | --- |
| 1 | What an Agent Is and When to Use One | 16% | ~8 | [D1](/openai/agents/domains/d1-what-an-agent-is/) |
| 2 | Defining Objectives and Tasks | **18%** | ~9 | [D2](/openai/agents/domains/d2-defining-objectives-and-tasks/) |
| 3 | Context, Tools and Permissions | **18%** | ~9 | [D3](/openai/agents/domains/d3-context-tools-and-permissions/) |
| 4 | Boundaries and Guardrails | 16% | ~8 | [D4](/openai/agents/domains/d4-boundaries-and-guardrails/) |
| 5 | Reviewing and Verifying Agent Work | 16% | ~8 | [D5](/openai/agents/domains/d5-reviewing-and-verifying-agent-work/) |
| 6 | Reliability and Iteration | 16% | ~8 | [D6](/openai/agents/domains/d6-reliability-and-iteration/) |
:::tip[Where the marks are]
Task Definition (18%) and Context and Tools (18%) together are **36%** of the material, and they are where beginners lose the most marks. Almost every failed delegation traces back to a vague objective or the wrong context and access — not to a weak model. Spend your revision time on D2 and D3 before polishing anything else.
:::
:::note[Two independent mock exams]
This track ships **two** full-length, domain-weighted independent mock exams. Mock exam 1 is your **diagnostic** — sit it untimed first to find your two weakest domains. Mock exam 2 is deliberately **harder** (more multi-constraint stems and `FIRST` / `BEST` / `MOST cost-effective` / `TWO` qualifiers) — use it as your **readiness gate** under timed conditions. All items across both mocks and the domain pages are distinct. Neither is an official OpenAI assessment.
:::
## The mindset this track rewards
Delegating to an agent is management, not prompting. The correct answers on this material consistently reflect the posture of a good manager handing work to a capable but literal new hire who will not ask for missing information and will not stop at a boundary you did not set:
- **Specify the outcome, not the keystrokes.** A good brief names the goal, the definition of done, the constraints and the sources of truth — then trusts the agent to find the path.
- **Give the least access that lets the work succeed.** More tools and broader permissions are more capability *and* more blast radius. Start narrow.
- **Put the gate before the irreversible step, not after.** An agent that can send, publish, pay or delete needs an approval checkpoint the *system* enforces, not one it merely promises to respect.
- **Verify the artefact, not the confidence.** You did not watch the run; trust the evidence trail and spot-checks, never a fluent summary of what the agent claims it did.
- **Treat a failed delegation as a brief defect first.** Before blaming the model, ask what the brief left ambiguous, what context was missing, and which checkpoint would have caught it.
Wrong answers reliably do the opposite: hand off an ambiguous goal, grant broad access "to be safe", trust a confident completion summary, and re-run the same vague brief hoping for a better result.
## Suggested time allocation
A focused plan of roughly 14 hours, weighted by domain weight and by how much judgment each domain demands.
| Domain | Weight | Hours |
| --- | --- | --- |
| Defining Objectives and Tasks | 18% | 3 |
| Context, Tools and Permissions | 18% | 3 |
| What an Agent Is and When to Use One | 16% | 2 |
| Boundaries and Guardrails | 16% | 2 |
| Reviewing and Verifying Agent Work | 16% | 2 |
| Reliability and Iteration | 16% | 2 |
## Hands-on preparation checklist
Do these in a real ChatGPT Work environment (or the closest surface your plan allows). Reading about delegation teaches you nothing about delegation.
- [ ] Take one recurring task you currently do by hand and write a **delegation brief** for it: goal, definition of done, constraints, sources of truth, checkpoints.
- [ ] Run the same task twice — once with a vague one-line prompt, once with the full brief — and compare how much rework each needs.
- [ ] Give an agent a task that requires a **connector or uploaded document**, then remove that context and watch where it guesses.
- [ ] Deliberately set an **approval checkpoint** before an action that leaves the workspace (an email draft, a shared file) and confirm the agent pauses.
- [ ] Delegate a multi-step task, then **verify the result without re-reading everything**: pick two claims and reproduce them from the evidence trail.
- [ ] Catch a **silent partial completion** — ask for five outputs, check whether you actually got five that meet the definition of done.
- [ ] Take a delegation that went wrong and **rewrite the brief once**, classifying the failure (misunderstood objective, missing context, wrong tool, partial completion, drift) before you change anything.
- [ ] Review a **long agent run** efficiently: read the plan, the checkpoints and the artefacts rather than the full transcript, and note what you would have missed.
## Track pages
16% · agent vs workflow vs prompt
18% · goals, done, constraints
18% · context and least privilege
16% · approval gates and limits
16% · evidence and spot-checks
16% · failure modes and iteration
50 items · diagnostic
50 harder items · readiness gate