Hand whole multi-step jobs to agents, supervise rather than type, and run two or three workstreams in parallel, with a test set and a permissions note.
Reporting season, a research task and a partner consolidation all land in the same fortnight, and each one is a sequence of steps you still run by hand.
You are one person doing three jobs. An agent that drafts the whole thing while you do something else is how the day gets longer.
A literature search across hundreds of candidates, or a first job where "look into this and come back with a brief" is the entire task.
Free knowledge, written for a non-technical reader. If you can explain each of these to a colleague in a sentence, you have the background this level needs.
Not useless, just not yet. Each of these appears at the level where it earns its place, or in a different course.
Print it, pin it next to the screen, and hand it to the colleague who asks how you do it.
| Item | Rule |
|---|---|
| Brief the job, not the step | Inputs · outputs · done when · stopping rules · ask me when · never do. |
| Blast radius first | What can this agent reach? What is the worst wrong action? If you cannot answer, do not start. |
| Read-only by default | Write access only where you will review every change. |
| Test set | Five cases with known answers. Run them before trusting anything. |
| Browser agent = you | It can do anything your logged-in session can. Log out of what it does not need. |
| MCP connection | What does it read? What can it write? Who approved it? Write it down. |
| Injection | A document or page can instruct the agent. Treat output built from untrusted sources as unverified. |
| Parallel means supervised | Two or three workstreams, each with a check-in point. Not ten. |
| Cost cap | Set a limit before a long run. Know how to stop it. |
| Judgement stays human | Scoring, awarding, hiring, paying, publishing: decisions, not steps. |
| Record | The brief, the permissions note, the test results, what was verified. |
Square brackets are yours to fill. Every example keeps personal data out and keeps the decision with the person.
Find foundations that funded [theme] in [region] between 2023 and 2026, from public sources only: annual reports, grant databases, official sites. Output a table: name, programme, typical grant size, last published call, source link. Stop at 25 candidates or when the sources run dry. Exclude anything without a primary source. Do not contact anyone. End with the list of what you excluded and why.Why it works: Inclusion criteria, a stopping rule, a source rule and a never-do, so the result can be checked rather than believed.
For each report in the folder Grantee reports 2026/Q3 (read-only), draft a board summary using the Board Summary Skill and save it to Drafts/. Stopping rules: if a report contains a named individual in a case story, do not summarise it; list it for me. If a figure has no source, write "not in source". Do not email anyone. When finished, give me the list of drafts and the list of skipped files.Why it works: A whole job with a Skill from level 4 inside it, read-only input, a personal-data stopping rule and no outbound action.
On the ministry's public tender portal, open each notice published this week in the category "water" and copy the title, deadline, reference number and link into a table. Do not log in, do not download attachments, do not submit anything. Stop after 30 notices. Tell me if any page asks you for information.Why it works: The agent works in a session with nothing to lose, with explicit never-do rules and a cap.
Here are twelve partner Excel files containing aggregate indicators only. Consolidate them into the indicator tracker format in this Project. Rules: flag any value that moved more than 30% from last month; never alter a partner's figure; if a file contains names or ID numbers, stop and tell me which file. Output: the consolidated table, the anomaly list, and a log of what you did.Why it works: Aggregate data only, an anomaly rule instead of silent acceptance, and a stop condition on personal data.
Search for peer-reviewed studies since 2018 on [topic] in [region]. Inclusion: empirical, in English or French, published in a journal or by a UN agency. Stop at 30 candidates. For each: citation, link, one-line finding. Then compare your list with the five papers I attach that I already know: did you find them? If not, explain why.Why it works: The five known papers are a built-in test set: if the run misses them, you know how much to trust the rest.
Before I trust this agent on the full season, here are five reports from last year and the summaries I wrote for them. Run the agent on these five only, then give me a side-by-side: where it matches, where it differs, where it invented something. Do not touch any other file.Why it works: Five known cases, compared before anything real is delegated. This is the proof of level 5 in one prompt.
Online or in person, in groups of up to nine, on your own tasks and documents. Every day ends with something that did not exist that morning.
Agentic AI shown on a live run. The blast-radius exercise on your own systems: what each tool can reach, what the worst wrong action would be. MCP connections and browser agents in your context, with a permissions note for each. Writing the job brief for one real job: inputs, outputs, stopping rules, ask-me-when, never-do. Permissions set read-only.
Build the five-case test set, run the agent on it, compare, fix the brief, run again. Then two workstreams in parallel with check-in points, a cost cap, and the verification of the result.
Thirty minutes is enough to place you on the ladder and say honestly whether this is the level to book. Or read level 6: Building next.