AI & Analytics Legends The knowledge platform for SAP Analytics
Concept card

SAP Joule Agent Builder

SAP Joule Agent Builder — Analytics Legends section illustration for the SAP Analytics knowledge base (concepts, studies, Academy)

As of 2026-10-10

What is SAP Joule Agent Builder?

The agent builder in Joule Studio (classic edition, GA Q4 2025) configures a Joule agent from expertise, instructions, a planning model and a response model, a thinking-step budget, and a small toolset (skills, document grounding, calculator, subagents, MCP servers), with environment-level masking and content filters.

The agent builder is the configuration surface in Joule Studio, classic edition, where you define a Joule agent: a component that receives a goal, plans, calls tools, reads the results and decides the next step until it can answer, asks the user, or runs out of its step budget. SAP announced in November 2025 that building custom agents in Joule Studio would be generally available in December, and the Q4 2025 release highlights list the agent builder as GA. It lives in SAP Build next to the skill builder (C161); the new, intent-based Joule Studio generates agents differently (C113).

What you configure

Per SAP Help ("Create a Joule Agent"), an agent is defined by:

Why it matters

  • The builder exposes the levers that drive quality and cost: planning model vs response model, thinking-step budget, toolset size and output format.
  • Environment-level orchestration profiles (masking, content filtering, prompt shield) and shared environments are the governance controls auditors will ask about.
  • Documented MCP constraints (streamable HTTPS only, no principal propagation, no interactive OAuth) decide which external tools are feasible.

Key points

  • Agent = description, expertise, instructions, additional context (##), base model + advanced model (+ backup), max thinking steps, tools, output format.
  • Tools: calculator, documents (SAP AI Core grounding on S3 or SharePoint), Joule skills (optional user confirmation), subagents; MCP servers on a separate tab.
  • Guardrails: environment-level orchestration profiles with data masking, Azure Content Safety filtering at each step and optional Prompt Shield.
  • Testing requires a private environment and the AI Agent Builder Activation toggle; the Timeline view exposes traces and request logs.
  • Vibe coding drafts configuration but cannot deploy, add MCP servers or change model settings.
  • Every quality/safety lever here (advanced+base model split, backup LLM, larger step budget) trades latency and AI Units cost for resilience or recovery ability — budget them for agents whose failure is expensive, skip them for low-stakes conversational agents.
  • MCP constraints for this builder (streamable HTTPS only, client credentials only, no principal propagation, no interactive OAuth) are a hard scoping filter: a server needing per-user isolation via principal propagation cannot connect here at all, regardless of tool design.
  • Environment orchestration profiles double as an evaluation signal: content-filter trigger rates and masking-entity frequency, tracked per agent, flag risky tool descriptions or instructions before an incident forces the review.
  • Joule Studio GA is 'coming soon' with no date (SAP Connect, Oct 2026): do not commit a client go-live to it; register builder agents in AI Agent Hub (GA) at creation.

Terms used on this page

Advanced model
The LLM in a Joule agent's model settings that formulates plans and selects the tools to use.
Orchestration configuration
Environment-level guardrail profile in Joule Studio that applies data masking and content filtering to deployed agents.
Prompt Shield
Optional input-filtering setting that detects prompt-injection attempts before they reach the agent.
Document grounding tool
Agent tool that retrieves from documents indexed through SAP AI Core grounding (AWS S3 or SharePoint sources).
Backup LLM
Optional secondary model configured in an agent's model settings, used automatically when the primary model is slow or failing.
Thinking-step budget
The maximum number of plan-act-observe iterations an agent may take before its task stops; validated between 1 and 100 in agent definitions.
sap-joule-studio-mcp-server property
The BTP destination property that must be set to true, with a streamable-HTTPS destination of the same name in both the Joule Studio and Joule subaccounts, to connect an MCP server to the agent builder.
Vibe coding (agent builder)
The agent builder's built-in AI assistant that drafts instructions, expertise and additional context from natural language, and can add or remove existing tools, but cannot deploy or change model settings.

Sources

  1. SAP Help — Create a Joule Agent (classic edition)
  2. SAP Help — Add a Tool (calculator, documents, Joule skill, Joule agent / subagents)
  3. SAP Help — Add MCP Servers to Your Joule Agent
  4. SAP Help — Orchestration Configuration for Your Agent (data masking, input filtering, prompt shield)
  5. SAP Help — Test Your Joule Agent
  6. SAP Help — Configuring Your Agent's Tools and Resources (best practices)
  7. SAP Help — Governance and Safety Controls (classic edition best practices)
  8. SAP Help — Optimizing Performance and Security (classic edition best practices)
  9. SAP Help — Creating Agents with Vibe Coding: Capabilities and Limitations
  10. SAP Help — Create a Document Grounding Tool
  11. SAP Help — Agent Configuration Properties (content-based agents)
  12. SAP News — New agentic capabilities on SAP BTP (5 Nov 2025: agent builder GA in December, A2A and MCP support)
  13. SAP News — SAP Business AI release highlights Q4 2025 (agent builder GA, AI Agent Hub GA)
  14. SAP News Center - Joule Work and the SAP Business AI Platform (8 Oct 2026)
  15. SAP Community - Joule Studio unlocked: From context to control - Joule Studio and SAP BTP for governed AI agents (Oct 6, 2026)
  16. SAP News Center - SAP Connect keynote: autonomous enterprise in action (7 Oct 2026)

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →