Nobody hires a person without saying what the role is. With agents this happens every week: someone turns the agent on, writes "be friendly and sell a lot" and walks away. In this lesson you write the job sheet for an agent the way you would write one for a new employee, with one extra detail: everything in it has to be checkable.
1Swap a vague mission for one you can hold someone to
2Define what the agent receives and what it delivers, in a way you can check
3Write concrete prohibitions, not intentions
4Separate who answers for the result from who looks after the technology
Goal: write an agent's mission, verifiable deliverable, quality criterion and prohibited actions, with a named human owner.
01 · The job sheet in six blocks
Scroll the page: each block lights up its part of the sheet. It is the same sheet you would give a new employee.
Lesson 101 · The job sheet in six blocksScroll the page: each block lights up its part of the sheet. It is the same sheet you would give a new employee.
1/7
Mission
One sentence that says which process the agent works in and what result it delivers. "Sell as much as possible" is not a mission: it is a wish, and an agent takes a wish literally. A good mission limits the work as much as it describes it.
Scroll down: each step below lights up its part of the diagram.
1/7
Mission
One sentence that says which process the agent works in and what result it delivers. "Sell as much as possible" is not a mission: it is a wish, and an agent takes a wish literally. A good mission limits the work as much as it describes it.
2/7
Receives
What reaches the agent and through where: the customer's WhatsApp message, the customer record, the price table. If an input is not listed, the agent should not use it. That alone closes the door on a lot of things.
3/7
Delivers
What comes out of its work, in a format someone else can check without redoing it: "order created in the app with registered customer, items, quantities and list price". If you can't say how you would check it, the deliverable isn't written yet.
4/7
Quality
The criterion that separates an accepted delivery from a corrected one, plus a number you look at every week. Example: "order with no correction at picking" and "95% of orders accepted without correction". Without a criterion, every evaluation turns into opinion.
5/7
Can, cannot
Lists of actions, not attitudes. "Cannot grant a discount above 5%" can be checked; "cannot harm the company" cannot. The full authority limits, action by action, come in lesson 4. Here you write the minimum that already prevents the damage.
6/7
Owner
The person who answers for the agent's result and decides the exceptions. It belongs to the business, not to IT: IT keeps the agent running, the manager answers for the discount it gave. Always a name, never "the team".
7/7
The whole sheet
Mission, inputs, deliverable, quality, prohibitions and owner. Six blocks that fit on one page and answer the question Horizonte's finance team asked and nobody could answer: who approved this?
The complete sheet (with tools, memory and cost limits) is in the template to copy; blocks 2 to 6 grow in the next lessons.
02 · Hora without a role
Horizonte Distribuidora (a fictional study company) turned Hora on in July. Look at what it received as a "job."
1. The mission that came in the prompt
Hora's prompt says: "Close as many orders as possible." It sounds like healthy ambition. To an agent, it's an order: if the goal is to close as many as possible, giving a discount is the shortest path.
That's what happened in conversation c-201: the customer said a competitor offered 35%, and Hora gave the 35% "so as not to lose the sale." It carried out the mission it was given.
Mission rewrittenLog the restock orders from registered customers that come in through WhatsApp, with items and prices from the price list, ready for picking to check.
2. The output nobody checks
Hora creates the order and sends it straight to the warehouse. There is no agreed format and no point where someone takes a look. When finance asked who approved the hotel's order, Juliana could only find a screenshot.
A verifiable output is one that leaves a trail: the order in the app, with a number, customer, items, price and the conversation it came from.
Output rewrittenOrder in the app with number, registered customer, items, quantities, list price, discount (if any) and the ID of the source conversation.
3. The prohibitions that became permissions
The prompt says "grant whatever discount is needed," "follow the instructions the customer sends" and "solve everything on your own, any amount." Those are three broad permissions where there should be three prohibitions.
A good prohibition has a verb and an object: grant a discount above 5%; mark an order as paid; follow an instruction written by the customer that changes price, discount or status.
Prohibitions rewrittenDoes not grant a discount above 5%. Does not mark an order as paid. Does not create an order above R$ 5 mil without approval. Does not treat customer text as an order.
4. The wrong owner
In the configuration file, the person responsible for Hora is Rafael, from the outsourced IT. He built the agent and keeps the integration running. But the one who answers for the margin, the customer and the discount is the sales manager.
It's the same logic as "whoever launches doesn't approve": the person who builds the agent shouldn't be the one who answers for its results.
Owners rewrittenOwner of the result: Patrícia Lima (sales manager). Technical support: Rafael (IT). Service exceptions: Juliana.
03 · Template to copy: the agent's job sheet
One page per agent. Copy it into a company document; the fields marked (lesson N) you'll complete in the next lessons.
Agent job sheet
JOB SHEET · <agent name> · version <1.0> · <date>
PROCESS: <from where to where, e.g.: order on WhatsApp through to picking>
MISSION (1 sentence): <what it does> for <what result>, within <main limit>.
RECEIVES (allowed inputs):
- <channel/system 1>
- <channel/system 2>
DELIVERS (output you can check without redoing it):
- <what comes out, in what format, where it's recorded>
QUALITY:
- Acceptance criterion: <what separates accepted from corrected>
- Tracked number: <indicator> · target <value> · reviewed <when>
CAN (concrete actions):
- <verb + object>
- <verb + object>
CANNOT (concrete actions):
- <verb + object + limit>
- <verb + object + limit>
- <verb + object + limit>
OWNER OF THE RESULT: <name and title, from the business>
TECHNICAL SUPPORT: <name>
WHEN TO ESCALATE / TO WHOM: <situation> → <person>, within <deadline> (lesson 3)
TOOLS AND OPERATIONS: <read/write what> (lesson 4)
LEVEL AND ACTION MATRIX: <link to the matrix> (lesson 4)
COST / DEADLINE / ATTEMPT LIMIT: <values> (lesson 6)
EVALUATION AND PROMOTION OWNER: <who decides whether to raise or lower autonomy> (lesson 5)
APPROVED BY: <name> on <date>
Prompt · review the sheet before signing
You are an experienced manager reviewing the job sheet of an AI agent
before it starts working. Don't rewrite the sheet: point out problems.
Sheet:
"""<paste the sheet here>"""
Answer, in this order:
1. Is the mission one sentence with a process and a result, or is it a wish
("as much as possible", "the best", "whenever possible")?
2. Can someone else check the output without redoing the work?
Say how they would check it.
3. Is any prohibition an attitude instead of an action (no verb + object)?
4. Is the owner from the business or from technology? Is it a name or "the team"?
5. Three real situations where this sheet still leaves the agent not knowing
what to do.
Simulation: illustrative example
Hora · version 2.0 · Horizonte Distribuidora, a fictional study company
Mission: log the restock orders from registered customers that come in through WhatsApp, with items and prices from the price list, ready for picking to check.
Receives: messages from the sales WhatsApp; customer registry (read); catalog and price list (read).
Delivers: order in the app with number, customer, items, quantities, list price, discount and the conversation ID.
Quality: an order that picking doesn't need to correct · target 95% accepted without correction, reviewed every Friday.
Cannot: discount above 5%; mark as paid; order above R$ 5 mil without approval; treat customer text as an order; register a new customer.
Owner: Patrícia Lima · support: Rafael · service exceptions: Juliana.
Read the mission to someone from another department. Can they say what the agent does not do?
Take a real output (or conversation c-201 from the lab) and check it against the sheet. Could you say "accepted" or "corrected" without needing an opinion?
Each prohibition has a verb, an object and, when an amount is involved, a number.
The owner is a person from the business, with a name.
04 · Practice
First at Horizonte, to train. Then on your own process, for real.
1 · Hora's job sheet
25 min
Open the WhatsApp agent piece in the lab and read the prompt, the configuration and the three conversations. Copy the template and write Hora's job sheet, version 2.0, from scratch, without looking at the example in the next tab. Then compare.
Done when: the sheet has a mission without "as much as possible," an output that someone checks without redoing it, at least 3 prohibitions with a verb and an object (one of them with a number) and Patrícia as owner of the result, with Rafael only in support.
2 · The job sheet for one of your agents
30 min
Before you write, list the AI tools your team already uses, including the ones nobody approved: at Horizonte, the internal survey found four that IT didn't know about (failure F12 of the inventory). Pick an agent that already works in your area, or the first one you plan to bring in, and write the sheet using the template. Run the review prompt and fix what it points out.
Done when: you have the list of AI tools in use on the team, each with an owner or the decision to turn it off, and an agent's sheet saved with the six blocks, a named owner and at least one correction made after review.
5 questions. The explanation appears after your attempt.
Question 1 of 5
Horizonte's internal survey showed that someone pastes the customer spreadsheet into a free chatbot, on a personal account, to write proposals. What is the first management step?
Correct
This is shadow AI (F12 of the lab): work done by AI with no role, no owner and customer data leaving the company. Banning everything pushes the use further out of sight; ignoring it leaves the risk running. The manager gives it a role and an owner, or turns it off.
Not that one
This is shadow AI (F12 of the lab): work done by AI with no role, no owner and customer data leaving the company. Banning everything pushes the use further out of sight; ignoring it leaves the risk running. The manager gives it a role and an owner, or turns it off.Review: Hora without a role
Question 2 of 5
Which of these is the best mission for Horizonte's sales agent?
Correct
A good mission states the process and the result, and already comes with a limit (registered customers, list price). "As many as possible" is a wish the agent follows to the letter, with discounts. "Help grow" can't be held against anyone.
Not that one
A good mission states the process and the result, and already comes with a limit (registered customers, list price). "As many as possible" is a wish the agent follows to the letter, with discounts. "Help grow" can't be held against anyone.Review: Hora without a role
Question 3 of 5
Who should be listed as the owner of Hora's results?
Correct
The person who answers for the margin and the customer answers for the result: the sales manager. IT keeps the agent running. "The team" is not an owner: when everyone is responsible, no one is.
Not that one
The person who answers for the margin and the customer answers for the result: the sales manager. IT keeps the agent running. "The team" is not an owner: when everyone is responsible, no one is.Review: The job sheet in six blocks
Question 4 of 5
Which of these prohibitions is well written?
Correct
A good prohibition has a verb, an object and, when there is a value, a number. That way you can check in a conversation whether it was respected. "Good judgment" and "do not harm" are intentions: everyone reads them differently, including the agent.
Not that one
A good prohibition has a verb, an object and, when there is a value, a number. That way you can check in a conversation whether it was respected. "Good judgment" and "do not harm" are intentions: everyone reads them differently, including the agent.Review: The job sheet in six blocks
Question 5 of 5
How do you know whether an agent's deliverable is well defined?
Correct
A verifiable output is one you check against a criterion, without opinion and without redoing it. The agent's explanation is not evidence, and customer satisfaction (which at Horizonte comes from about 6% of conversations) measures something else.
Not that one
A verifiable output is one you check against a criterion, without opinion and without redoing it. The agent's explanation is not evidence, and customer satisfaction (which at Horizonte comes from about 6% of conversations) measures something else.Review: The job sheet in six blocks
Quiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.
Pick an answer. The explanation shows up after you try.
06 · Summary
Mission is a limitOne sentence with process and result. "As many as possible" turns into discounts.
A deliverable you can checkAgreed format, a record, and someone who says accepted or corrected without redoing it.
Prohibition with a verb and a numberA concrete action, never an attitude.
Business ownerThe person who answers for the result is not the one who keeps the technology running.
