Joule Agents in Production — Testing and Lifecycle
As of 2026-09-25
For SAP AI consultants moving from building a first Joule agent (M326) to operating one at scale. Grounds a production delivery plan in SAP AI Agent Hub's five documented lifecycle stages (Plan and Build, Discover and Provision, Observe and Analyze, Secure and Govern, Optimize and Decommission); the new Joule Studio's three-part testing model (generated automated tests, AI-powered evaluation, Try mode) and its GitHub Actions approval gate; the six documented human-in-the-loop mechanisms, chosen by reversal cost; the four documented multi-agent coordination mechanisms and their hard limits (5 direct / 25 total subagents); observability signals (token usage, inference observability GA 8 May 2026); a custom evaluation harness (200-500 QA pairs) instead of BLEU/ROUGE; and the EU AI Act Article 15 enforcement date (2027-12-02) for high-risk Joule use cases. Three exercises and a self-assessment close the module.
What you will learn
- Structure an agent delivery plan around SAP AI Agent Hub's five documented lifecycle stages instead of an ad hoc build-test-deploy checklist
- Explain the three-part testing model of the new Joule Studio (generated automated tests, AI-powered evaluation, Try mode) and the GitHub Actions approval gate for production deployment
- Select the right human-in-the-loop mechanism among the six SAP documents, based on how costly the action is to reverse
- Choose among the four documented multi-agent coordination mechanisms (Joule Assistants, classic-builder subagents, agent-request, A2A) and respect their hard limits, notably the 5-direct/25-total subagent ceiling
- Design a custom evaluation harness (200-500 question-answer pairs, red-teaming) instead of relying on BLEU/ROUGE for factual SAP tasks
- State the EU AI Act Article 15 enforcement date (2027-12-02, Digital Omnibus) and its consequence for high-risk Joule use cases
Module overview
Who this is for. You built your first Joule agent in M326 and know what the orchestration pipeline does (M325). The question the client asks has changed: not "can Joule do this," but "who owns it when it misfires at 2 a.m., who signs off before it posts a credit memo, and how do we prove to an auditor it still behaves the way it did on day one." This module answers that second question — keeping an agent alive, tested and defensible in production — using the lifecycle, testing and governance mechanisms SAP documents for the new Joule Studio and SAP AI Agent Hub as of September 2026, not a generic MLOps framework relabelled for SAP.
Prerequisites
- M326 — Building Joule Agents with Joule Studio, and M325 — SAP Generative AI Hub hands-on
- Comfort with the lifecycle vocabulary of module M333 (evaluation, RAG, tool calling) and with reading a governance workflow diagram
- Optional: access to a Joule Studio or SAP Build trial tenant for a hands-on E1/E2
Outcomes
- Write a statement of work for an agent delivery that names all five AI Agent Hub lifecycle stages and their owners.
- Justify, for any given agent action, which of the six human-in-the-loop mechanisms applies and why.
- Design a coordination pattern for a multi-agent scenario using one of the four documented SAP mechanisms, respecting subagent limits.
- Produce an evaluation-harness outline that would survive both an internal quality review and an EU AI Act Article 15 audit.
Full module available to members. The full module adds: the decision framework · the end-to-end scenario walkthrough · the KPI scorecard · the anti-patterns · the code blocks · the knowledge check · the diagrams.