Agent Design Patterns for SAP Processes — Tools, Planning, Human in the Loop
As of 2026-09-25
The synthesis module of the SAP AI batch: a repeatable method for designing an agent for an SAP process, built around four elements (expertise/instructions, tools, planning, human checkpoints) that map directly onto SAP's own Joule Studio agent builder (planning model, response model, thinking-step budget, a small toolset of skills, document grounding, calculator, subagents and MCP servers). It covers composing agents through subagents, A2A (Agent Card, Task, Message, GA alongside the agent builder in Q4 2025) and the early-preview n8n integration in Joule Studio, gives a design checklist derived from EU AI Act Article 14's five human-oversight capabilities, and closes the loop opened by M370 (Signavio recommends an agent) and M371 (the AI Agent Hub governs it) and supplied by M372 (CAP/MCP tools).
What you will learn
- Name the four elements every agent design needs (expertise/instructions, tools, planning, human checkpoints) and where SAP's Joule Studio agent builder implements each
- Classify a tool by risk (skill, document grounding, calculator, subagent, MCP server) and state who is accountable for its worst-case failure
- Explain a planning loop's thinking-step budget as a safety control, not only a cost control, and design its failure path
- Choose between a subagent, an MCP tool and an A2A delegation for composing agents, and explain the A2A Agent Card / Task / Message model
- Map EU AI Act Article 14's five human-oversight capabilities onto concrete agent design mechanisms
- Trace an agent design through to AI Agent Hub registration (M371) and, where it exposes a CAP/MCP tool, an SAP API Policy check (M372)
Who this is for. M333 taught you to call a model and let it call a tool. This module is about the step after that: designing something that keeps acting — reading state, choosing a tool, observing the result, deciding what to do next — until a goal is met or a human has to step in. That is an agent, and SAP gives you a concrete, documented way to build one (Joule Studio) plus two protocols for wiring it to tools and to other agents (MCP, A2A). This module gives you a repeatable design method: four elements every agent design needs, where SAP's own tooling implements each one, and where the human checkpoint belongs — closing the loop this batch opened with M370 (Signavio recommends an agent), M371 (the Agent Hub governs it) and M372 (CAP/MCP can supply its tools).
1. Four elements every agent design needs
Whatever framework builds it, a usable agent design answers four questions, and SAP's own agent builder in Joule Studio asks them explicitly rather than leaving them implicit in code:
Prerequisites
- Completion of M333 (AI & LLM Fundamentals — tool calling, agents, MCP, A2A) and strongly recommended M372 (CAP and the SAP Cloud SDK for AI)
- Recommended: M370 (AI in SAP Signavio) and M371 (AI in SAP LeanIX), since this module closes the agent lifecycle both begin
- Comfort reasoning about risk and accountability for an automated system's actions, not just its correctness
Outcomes
- Design an agent's toolset scoped to the smallest set that does the job, with a named accountability for each tool's worst-case failure.
- Set and justify a thinking-step budget with an explicit failure path (partial result, escalation, or clean failure).
- Choose correctly between subagent, MCP tool and A2A delegation when composing multiple agents for one SAP process.
- Produce a five-line EU AI Act Article 14 mapping for any agent design, naming the exact mechanism behind each oversight capability.
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.