BTP & Integrationbeginner

Joule Studio Part 2: Build Your First Custom Joule Agent

Key takeaway

Walk through building a Joule agent end to end: prerequisites (BTP entitlements, destinations), defining the agent's goal and instructions, adding a skill as a tool, testing in the studio, and deploying so it's reachable from Joule. Principal propagation carries the calling user's identity through every skill and service call automatically.

Explanation

📌 KEY POINTS
✅ Prerequisites: the right BTP entitlements/subaccount setup, and destinations configured for any back-end system the agent's skills will call.
✅ An agent is defined by a goal/instructions (what it should accomplish and how) plus a set of tools — typically skills you or others have built.
✅ Principal propagation is enforced automatically: the calling user's identity passes through to every skill and CAP service call, so back-end authorizations apply per user, not per shared technical account.
✅ Test the agent inside Joule Studio before deploying — conversational testing surfaces gaps in instructions early.
⚠️ An agent with vague or overly broad instructions will improvise in unpredictable ways; specific, bounded instructions produce reliable behavior.

━━━━━━━━━━━━━━━━━━━━
▶ WHERE WE ARE
Part 1 explained the landscape. Now building.

━━━━━━━━━━━━━━━━━━━━
▶ 1. PREREQUISITES

BTP entitlements, destinations.

━━━━━━━━━━━━━━━━━━━━
▶ 2. GOAL AND INSTRUCTIONS

What the agent should do, and its boundaries.

━━━━━━━━━━━━━━━━━━━━
▶ 3. TOOLS

Skills (and other agents) the agent can call.

━━━━━━━━━━━━━━━━━━━━
▶ 4. PRINCIPAL PROPAGATION

Automatic, per-user identity through the call chain.

━━━━━━━━━━━━━━━━━━━━
▶ 5. TEST AND DEPLOY

Conversational testing in studio, then deployment.

━━━━━━━━━━━━━━━━━━━━
▶ STEP BY STEP

Configure destinations for the back-end systems involved.
Define the agent's goal and add one skill as a tool.
Test conversationally in the studio, then deploy.

✅ CHECKPOINT: a test user with limited authorizations gets correctly limited results from the agent.

Code example

ABAP Code

===== AGENT DEFINITION (conceptual) =====Goal: "Help employees check the status of their IT support tickets and create new ones."Instructions: "Only discuss tickets belonging to the requesting user. If asked about another user's ticket, decline and explain."Tools: [GetTicketStatusSkill, CreateTicketSkill] ===== IDENTITY FLOW =====User asks Joule → Agent invoked → Skill called → CAP service called  (user's identity propagated automatically at every step — no shared technical user) ===== TEST THEN DEPLOY =====Test in studio with a real, limited-privilege test user → verify authorization boundaries → Deploy

Real project scenario

A team built an agent with the instruction "help with HR questions" and no boundaries; it began answering questions about other employees' compensation because a poorly scoped skill technically allowed the lookup. Tightening both the instructions and the underlying skill's authorization fixed it — a reminder that instructions alone aren't a security boundary.

Common mistakes

Vague, unbounded instructions.
Testing only with admin/privileged accounts.
Assuming instructions alone enforce security (they don't — the underlying skills and back-end authorizations do).
Skipping destination configuration until deployment fails.
Deploying without conversational testing.

Best practices

Write specific, bounded instructions.
Test with realistically-scoped users.
Enforce security in skills/back-end authorizations, not just instructions.
Configure destinations early.
Test conversationally before deploying.

Interview angle

Questions: "What are the prerequisites for building a Joule agent?", "How does principal propagation work in Joule Studio?", "Why isn't an instruction like 'only show the user's own data' sufficient on its own for security?"
Strong answer: prerequisites, goal/instructions/tools structure, automatic principal propagation, the limits of instructions as a security control, test-then-deploy discipline.

Practice this lesson

Quick check

1. Why is a human approval step important for an agent proposing a consequential action?

2. Why does an agent need explicit action boundaries in its instructions?

3. A maintenance agent could identify missing spare parts but had no way to actually order them. What is missing?

Interview questions

1. A production issue is reported in Joule Studio and AI Agents around Joule Studio and AI Agents: Beginner Concepts. How would you troubleshoot it systematically and avoid the common failure patterns?

Answer it in your own words first, then compare.

2. You are the senior consultant/architect responsible for Advanced Joule Studio and AI Agents: Architecture, Integration and Production Design. The project situation is: A maintenance agent could identify missing spare parts but originally had permission to create purchase orders directly. How would you assess the design, identify the highest-risk failure modes, and define the evidence required before approving it for production?