Now accepting guest editorial from SAP, Oracle, Workday and ServiceNow practitioners.Pitch a guide
The Sign-Off Table

Lesson 2 of 6

Designing an agent: goal, instructions, tools, boundaries

10 min

Listen to this lesson

0:00 / 8:01

Most weak agents are weak on paper before anyone opens the builder. Elena and Maria spend a morning writing the design: what the agent is for, what it must never do, which tools it gets, and when it hands over to a person. SAP's best practices for Joule agents say the same thing in builder terms: be clear and specific, use step-by-step logic, avoid conflicting rules, and restrict the toolset to what the job needs.

The agent design canvas

ElementQuestionFleet Order Status agent
GoalWhat outcome, for whom?Give O2C clerks a correct status for a fleet customer's order in one answer
UsersWho may use it?Maria's O2C team in sales area 1710 / 10 / 00; not customers directly
InputsWhat does a request contain?Sales order number, customer PO number (e.g. GLF-PO-55120) or customer name
ToolsWhat may it read or change?Read: order, delivery, billing status. Change: none directly; may start an expedite request process
KnowledgeWhat documents ground it?Contoso's fleet service policy and shipping cut-off times
BoundariesWhat must it refuse?Price changes, credit decisions, anything for company code 1910
EscalationWhen does a human take over?Credit block, missing order, customer complaint, or two failed lookups
SuccessHow will we know it works?Evaluation set accuracy and time saved (lesson 5)

Write the canvas before you write instructions; each row becomes a field, a tool, or a test case.

From canvas to the agent editor

In the classic agent builder you choose Create > Joule Agent, give a name and a description, and fill three text fields. SAP's tutorial explains that the description helps Joule decide when to activate the agent, so it must be specific and business-oriented. Expertise is a short summary of what the agent is good at. Instructions are the step-by-step behavior: what to collect first, how to evaluate results, how to conclude. Additional Context carries tone and rules. Together with the tool descriptions, these four elements form the agent's prompt, as course sap-ai-201 explained.

The intended behavior, written as a flow before it becomes instructions. Each branch becomes an evaluation case.
The Sign-Off Table · thesignofftable.com/academy/sap-ai-207
  1. Expertise: "You help Contoso Motor Co. order-to-cash clerks find the current status of fleet customer sales orders in sales area 1710 / 10 / 00."
  2. Instructions: "1. Identify the sales order from the order number or the customer purchase order number. 2. Retrieve order status. 3. If the order has a credit or delivery block, stop and escalate. 4. Otherwise retrieve delivery and billing status. 5. Answer with document numbers, quantities and dates. 6. Offer an expedite request only if delivery is not yet goods-issued."
  3. Additional Context: "Be brief and factual. Never promise delivery dates the documents do not show. Do not discuss prices, discounts or credit limits. If a request concerns company code 1910, say it is out of scope."

Tools and boundaries

The toolset is the hard boundary. An agent can only do what its tools allow, however its instructions are phrased, so Elena gives it read skills for status and one process skill that requests an expedite rather than changing the delivery. Write-capable tools need a confirmation step: SAP's skill guidance calls a confirmation before any create, update or delete a critical requirement. Keep tool descriptions precise, because the model chooses tools from their descriptions; two tools with overlapping descriptions produce random choices.

Let the agent actRead-only status lookups, drafting a message for review
Act with confirmationStarting an expedite request that a person approves
Recommend onlySuggesting a credit review; the clerk decides
Keep out of the agentPrice changes, credit releases, cancellations
Place each candidate tool before you add it: the matrix decides between act, confirm, recommend or exclude.
The Sign-Off Table · thesignofftable.com/academy/sap-ai-207

Designing the handover

Escalation is a feature, not a failure. Write down who receives the handover, through which channel, and what the agent passes on so the person does not start from zero: the order number, what the agent already looked up, the reason for escalation and the clerk’s question in the clerk’s words. For Contoso, credit blocks go to the credit team’s queue and complaints go to the O2C supervisor. The agent then tells the clerk who has the case and what happens next. Test the handover content as carefully as the answers: an escalation with missing details simply moves the work to someone else.

A designed escalation passes the context on and tells the clerk who owns the case now.
The Sign-Off Table · thesignofftable.com/academy/sap-ai-207

What would you do?

Maria wants the agent to release credit blocks for Great Lakes Fleet Services when the clerk asks, "because the customer always pays." What should Elena do?

Key points
  • Write a design canvas: goal, users, inputs, tools, knowledge, boundaries, escalation, success.
  • The agent editor's description, Expertise, Instructions and Additional Context plus tool descriptions form the prompt.
  • The toolset is the real boundary: restrict it; write actions need confirmation; high-exposure actions stay out.
  • Every branch of intended behavior becomes an evaluation case.

Check your understanding · not graded

Lesson check

  1. 1. Why does the agent description matter in the classic agent builder?

  2. 2. What is the most reliable way to stop an agent from changing prices?

  3. 3. Which field holds the step-by-step behavior the agent should follow?

  4. 4. Two tools have overlapping descriptions. What problem is most likely?