BlogBlogA post on somebody's own site or a community platform. Nobody edits it after publication, so it dates from the day it was written.See every one →SAP Community
Building an LLM Agent using CAP
The argument is that CAP is already most of what an agent framework wants — services, handlers, a typed model — so the agent is built inside it rather than beside it. AI Core through sap-ai-sdk, then Langchain for the agent and its tools. Long, with a TLDR at the top for people who want the shape without the walk.
Where it sits
Explains
Agent toolingCAPBTPAgent toolingBuilding an agent into your own application — the tools it may call, the state it keeps between turns, and the framework it runs in. Distinct from MCP servers, which expose an application to somebody else's agent, and from agent guidance, which is knowledge handed to a coding agent.LLM integrationCAPBTPLLM integrationCalling a language model from application code with the tenant, destination and credential handling already solved instead using an SDK call with a hand-rolled key.
Since then
Agent tooling changed in @cap-js/agents 0.9.1, 2026-08-14
Source
reachable
Who is behind it
Johannes VogtJohannes VogtWrote this
In the atlas: maintains 1, contributes to 28, wrote 10@Akatuoroauthor
In the atlas: maintains 1, contributes to 28, wrote 10@Akatuoroauthor
Before you read it
read from the pageApplicabilityCore changedAgent tooling changed in @cap-js/agents 0.9.1, 2026-08-14.
This is a derived prompt to re-check the work, not a claim that the work is wrong. Exact stale statements are recorded separately after editorial review.
Properties
About this work
What was found on the page
Prose
2,537 words
Code blocks
40
Repository links
1
For agent-tooling
Applies in
For llm-integration
Applies in
Group by71 of 71 pieces
Piece, what it offersStill applies
Building Code-Based AI Agents on the BAIPBuilding Code-Based AI Agents on the BAIPA Devtoberfest 2026 session on writing the agent yourself rather than configuring one, which is the agent tooling edge: an agent is a model with tools, skills, memory and context around it, and the session reads the code of one built in CAP as a LangGraph state graph, with an agent node that calls the model, a tools node holding the MCP clients, and the edge between them looping until no tool is needed. It then shows the shorter route, a CDS service annotated for the agent protocol with its instructions and skills in an agent.md file. SAP AI Core carries both, with the CAP service in front of it as middleware so the credentials never reach the browser, and that is what cap-agent-ui5-webui is demonstrated against: added as a dependency, it gives the annotated service a UI5 chat interface that can be deployed rather than only previewed. A Python multi-agent application on Cloud Foundry, where an orchestrator delegates to per-stack specialists so no single model sees every tool, closes the session.published 2026-10-09
CAP in the age of AI #5 - agents wrap-upCAP in the age of AI #5 - agents wrap-upThe agents topic closed out on what sits around the agent rather than inside it, now that the documentation has moved into capire. The chat preview's session is persisted in the database rather than kept in memory, traces go out through @cap-js/telemetry to a self-hosted MLflow under the OpenTelemetry conventions for generative AI, and the panel is explicit that @agent reuses the MCP adapter's code programmatically without serving an MCP endpoint. It ends in the xtravels sample, where a travel agent delegates to a hotel-service sub-agent and reaches a separately started flight service over MCP through the service binding CAP resolves anyway. Earlier in the hour the same agent answers from a model running on the presenter's laptop, reached as an OpenAI-compatible endpoint.published 2026-10-02
What's in Daniel Hutzel's brain?What's in Daniel Hutzel's brain?A Devtoberfest 2026 session in the building-apps track, in which Daniel Hutzel and DJ Adams work through what the CAP team is building for agents. A bookshop gets an MCP endpoint beside its OData one from a single annotation, with the plugin writing the coding agent's configuration itself; in the xt-travels sample five services that know nothing about each other are served over MCP and an agent does the orchestration the hand-built UI and its value helps used to do. Hutzel then annotates a service as an agent so the ReAct loop and the choice of model move into the CAP service, and finally drops MCP for a skill that teaches the agent CQL instead, which is where the edges to Serve your application over MCP, Agent tooling and Agent guidance come from. The mechanism under all of it, a plugin file in the package root and event handlers that can be added to any CAP service, is the edge to Plugin authoring.published 2026-10-02
Real Pro-Code Development with Joule StudioReal Pro-Code Development with Joule StudioA Devtoberfest 2026 AI-track session that runs one agent build across both of Joule Studio's paths. Playing a procurement manager, one presenter states the intent in the browser, answers the clarifying questions, watches the fit-gap analysis name the S/4HANA APIs the use case needs, and pushes the intent and requirements document to GitHub; the other clones it and carries on in a terminal, logging in with the jl command line against her landscape URL and running its init command to fetch Joule Studio's skills for OpenCode. From there the coding agent writes the specification, the discovered API specifications, the MCP translation and the agent's Python file with its model configuration and evaluation scenarios, which is the edge to Agent tooling; the skills themselves, which SAP says not to edit but which a developer may add to, are the edge to Agent guidance.SAP Developers, published 2026-10-01
🟣 Enable Your AI application to Run on Multiple Models to Ensure Resilience and Compliance🟣 Enable Your AI application to Run on Multiple Models to Ensure Resilience and ComplianceDevtoberfest 2026 session AI-21. A cross product owner on the SAP AI Core team gives the reasons an AI application ends up on more than one model — a model down at the hyperscaler, a data centre that is not allowed to run it, a version being deprecated — and runs a joke app that answers from Gemini 2.5 Flash in EU10 and from Claude 4.5 Haiku in Asia-Pacific. The walk through AI Core's parts is where the edges hold: prompts kept as templates in the prompt registry, imported through the launchpad and the APIs or synced as YAML from a git repository (Prompt management); the evaluation service scoring one prompt across up to four models per run on built-in or model-as-judge metrics, and the prompt optimizer generating variants against one metric (Agent evals); and the orchestration service harmonising the providers behind one message format with a priority-sorted fallback list (LLM integration). The session also states where a metric has to be designed rather than picked: an exact-JSON match scores an email generator at zero, so the joke app's custom metric judges topic and family-appropriateness instead.SAP Developers, published 2026-10-01
What an AI infra engineer sees when opening SAP's AI stack (1/3): the model gatewayWhat an AI infra engineer sees when opening SAP's AI stack (1/3): the model gatewayA first look at SAP AI Core's Generative AI Hub from someone who builds gateways outside SAP for a living: the orchestration service harmonizes provider calls onto one API (a model name string switches providers, and SAP Note 3437766 carries the current model list, token rates and deprecation dates), runs a fixed, non-reorderable module pipeline — templating, masking through SAP Data Privacy Integration, grounding, content filtering, then the model call — and retries down an ordered fallback list whose entries can each carry their own model and rules. It places @sap-ai-sdk/orchestration among the official JavaScript, Java and Python SDKs, and notes SAP's own LiteLLM provider as a way to put a different gateway in front of it.amishkush, published 2026-09-30
CAP: The Road Ahead - Simplicity in an AI EraCAP: The Road Ahead - Simplicity in an AI EraDevtoberfest 2026 session CAP10, from the head of SAP's Cloud Application Programming Model: forty minutes of live building rather than slides, on the argument that the cheapest code in an AI project is code nobody has to generate. Claude Code builds an invoice-matching workspace from cds init onwards, and each step is one the framework answers rather than the agent: a 42-line domain model that @sap/cds serves as OData and Fiori elements with a status flow, data privacy and change-tracking annotations that produce audit log entries in the console and would reach the SAP Audit Log service once deployed, an imported S/4HANA supplier data product, and semantic matching from a local embedding model after @cap-js/ai is added. The last three steps are annotations: @mcp beside @odata turns the service into an MCP server that a second Claude session queries through @cap-js/mcp, an agent annotation adds the A2A protocol with an agent card and an audit log entry for the agent task, and a skill file gives the agent the matching policy.SAP Developers, published 2026-09-25
Meet Joule Studio and Build your First AI AgentMeet Joule Studio and Build your First AI AgentA Devtoberfest 2026 session from the product manager of Joule Studio, walking one agent through the whole path the tool defines: an intent, the problem statement and goals a coding agent derives from it, the business context it takes from the SAP knowledge graph, and the solution type it proposes with a stated intent fit. The demo is a collections agent that drafts dunning emails: the product requirements document names the metrics and the human-in-the-loop review, the technical specification lists the artifacts to build — the MCP tools among them — and the build is followed by a business test run against business partner 10001, technical unit tests, and deployment to the managed runtime, where the agent and its MCP server appear as deployed assets. Two points shape how it is used rather than what it produces: every asset is a file, editable in the web editor or in the code view and pushable to git, and the same solution can be built from local pro-code tools; n8n workflows are stated as delivered out of the box inside the studio.SAP Developers, published 2026-09-24
Combining AI Agents with the power of SAP HANA embedded ML (part 1)Combining AI Agents with the power of SAP HANA embedded ML (part 1)Devtoberfest 2026 session DAT-12, the first of two parts, in which Vitaliy Rudnytskiy does a data scientist's work against the embedded machine learning of SAP HANA Cloud by talking to an agent. The session first takes an agent apart into its LLM, its tools, its memory and the loop between them, then stacks the Predictive Analysis Library under the hana-ml Python wrapper and that under the HANA AI toolkit, whose tools an agent calls from plain English in a notebook to preview a sales refunds table, split it into train and predict tables, run a time-series check and train an AutoML forecast, keeping what it decided in Markdown memory files. That is the agent tooling edge; the same toolkit is then started as an MCP server over stdio and over streamable HTTP and reached from a coding agent and from Joule Work's desktop, which is where the live demo partly fails and the failure is shown.published 2026-09-22
CAP in the age of AI #4 - more on agentsCAP in the age of AI #4 - more on agentsAsking a human first, and telling the agent how to behave. Ordering a book through @cap-js/agents happens without a question being asked; the human-in-the-loop annotation on the action turns the same request into an approve-or-reject step, which the episode follows down to the A2A status update that stores the task in an input-required state in the database — A2A itself being JSON-RPC 2.0 over HTTP with server-sent events for the stream. The rest is instructions on disk: an agents.md in the service's own folder removes the question about how many copies to order, a skill under a skills folder is advertised by its description and read only when the agent decides the case applies, and either file makes the plugin build a deep agent instead of the plain one.published 2026-09-11
SAP's API Policy on AI Agents: What It Prohibits, Allows, and Leaves OpenSAP's API Policy on AI Agents: What It Prohibits, Allows, and Leaves OpenReading SAP's API Policy v.4.2026a and its FAQ against the reference architectures, the piece narrows the widely repeated claim that SAP bans AI to what Section 2.2.2 actually restricts: an agent planning and chaining API calls outside an SAP-endorsed route, closed off further by Section 3's block on bypassing API Controls through custom code. It compares the endorsed and allowed paths — the MCP Gateway on SAP Integration Suite reaching S/4HANA and ECC through RFC-based back ends under a 15-tool, 200,000-token limit, SAP-provided MCP servers through Joule Studio, the A2A protocol, customer-operated MCP servers allowed under FAQ Question 56, and the ADT MCP Server, whose 18 tools work only on development artifacts and are documented as not intended for productive use. It leaves open whether Section 2.2.2 covers ABAP code that calls a model from inside the system, since enforcement is defined at the API-traffic level rather than by a list of tools.published 2026-08-31
UI5 AI Features library is now an npm PackageUI5 AI Features library is now an npm PackageThe release note for a demo turning into a dependency, and it is explicit about why that matters: the earlier version worked, but using it meant cloning the repo and copying both halves, which stops working the moment one of them changes. The post walks through what shipping the CAP plugin and the be.wl.ai control library in one npm version buys — a client and server contract that cannot drift — then lists the eleven controls by group, shows the two-line configuration that activates each half, and notes that this version is actually deployed on BTP behind both a standalone approuter and Work Zone. It closes on the demo repository, written with no library source in it as the test that the package stands alone.published 2026-08-31
UI5 AI Features library is now an npm Package
cds 8, line ended. Affects Runnable code.
Connecting SAP Integration Suite with SAP Document AI and SAP Business AIConnecting SAP Integration Suite with SAP Document AI and SAP Business AIWalks through a Cloud Integration flow in SAP Integration Suite that pulls PDF purchase orders from a Gmail mailbox, extracts header and line-item fields through SAP Document AI's job-submission and polling API, and validates the result with SAP AI Core's generative AI hub using GPT-5-mini as a quality check rather than the source of the extracted values. Groovy steps combine the Document AI confidence score and the model's score into one figure that routes low-scoring purchase orders to manual review, with no S/4HANA posting performed.subhojitSaha, published 2026-08-24
CAP in the age of AI #3 - from MCP to A2ACAP in the age of AI #3 - from MCP to A2AThe episode that finishes @cap-js/mcp with authorisation and then crosses to agents. @requires and @restrict are the same annotations CAP already has, and the plugin reads them to filter the tool list and the describe output per user and per request, so a model is not told about entities the caller cannot reach — where the OData metadata of the same service lists every exposed entity regardless. The second half annotates a service with @agent from @cap-js/agents, which serves it over A2A with an agent card and a chat preview, runs first against a mocked response and then against SAP AI Core bound with cds bind and a hybrid profile, and reuses the MCP adapter's tools rather than defining new ones.published 2026-08-21
Adapting SAP Joule for Purchase Order Creation from an Existing Purchase RequisitionAdapting SAP Joule for Purchase Order Creation from an Existing Purchase RequisitionA custom Joule Agent and Joule Skill, built in Joule Studio, classic edition to extend SAP Joule past its built-in Purchase Order display and change features: a buyer's single sentence naming a Purchase Requisition number becomes a created Purchase Order, with the skill wrapping the Purchase Requisition and Purchase Order OData V4 APIs as Action Packages reached through two destinations. The author anchors the agent's success condition to a PurchaseOrderNumber in the skill's output rather than an HTTP status code, and reports that the Purchase Requisition item structure does not map one-to-one onto the Purchase Order creation payload.jonathanvi59, published 2026-08-19
SAP Joule: Invoice Agent for Faster Supplier Invoice Release in SAP S/4HANA Accounts PayableSAP Joule: Invoice Agent for Faster Supplier Invoice Release in SAP S/4HANA Accounts PayableWalks through building an Invoice Release Agent in SAP Joule: a custom OData service surfaces supplier invoices whose S/4HANA Flexible Workflow approval is overdue, the agent ranks them by invoice value, aging, payment tenor or a combined score, and a Joule Skill calls the standard SAP AP Invoice OData API to release them, re-prioritizing the remainder after each action. It sets out the detect-prioritize-act loop and the four ranking criteria in enough detail to rebuild the pattern, and leaves authorization, auditability and segregation-of-duty controls to the reader's own design.idc, published 2026-08-11
CAP in the age of AICAP in the age of AISeven live episodes of Hands-on SAP Dev in which DJ Adams works through CAP's AI features with the people who build them, each one a screen-share against a sample project. The first three are @cap-js/mcp: the @mcp annotation on a service, the describe, query and call tools, CQL as the query format, and what authorisation takes out of the tool list. From episode three the run moves to A2A and @cap-js/agents — human approval of an action, agents.md and skills, session persistence, telemetry and sub-agents — and the last two episodes are vector embeddings in CDS, local on SQLite and native on SAP HANA Cloud.published 2026-08-07
Building Durable AI Agents - Announcing the cds-langgraph-persistence PluginBuilding Durable AI Agents - Announcing the cds-langgraph-persistence PluginThe announcement, and the argument behind it. Treating an agent as a software architecture rather than a model with tools makes a list of production requirements — identity, authn and authz, multitenancy, persistence, event-driven flows, scalability, observability, CI/CD — and the author's point is that CAP already answers most of them for BTP, which is what makes LangGraph inside CAP Node.js worth doing. The post then narrows to one of those requirements, durability, and the checkpointer that supplies it. Written from a year of building agents next to real business systems rather than as a framework tour.merbinjanselm, published 2026-07-15
Building Durable AI Agents - Announcing the cds-langgraph-persistence Plugin
@mi8y/cds-langgraph-persistence has been through 3 major lines since this was published: 0.5 (2026-07), 0.6 (2026-07), 0.7 (2026-09). Affects Runnable code.
Autonomous Diagnostics on Kyma Unplugged: Goose + LiteLLM + KRO + SAP AI Core in ProductionAutonomous Diagnostics on Kyma Unplugged: Goose + LiteLLM + KRO + SAP AI Core in ProductionAn in-cluster autonomous diagnostic pipeline on Kyma Unplugged (open-source Kyma, noted to also run on SAP BTP, Kyma runtime): a Goose agent with kubectl access calls SAP AI Core through a LiteLLM proxy, and most of the post is about the quirks of that route — drop_params silently swallowing unrecognised parameters and beta headers, streaming that has to be off when tools are attached, and prompt caching that sits on the Orchestration V2 API rather than the foundation-model path the LiteLLM provider hits. A second Node.js proxy reaches that API, and the author reports caching unconfirmed at production scale, which leaves the roughly $0.004 per static run a target rather than a measured figure. Spend metrics reach SAP Cloud Logging over OTLP through Kyma's Telemetry MetricPipeline while the health analysis itself lands in Kubernetes custom resources read in Busola, and the stack is declared as KRO Resource Group Definitions applied with one Terraform command.quovadis, published 2026-07-09
Natural Language Queries for SAP CAP via MCP: The LLM Translates, CDS ExecutesNatural Language Queries for SAP CAP via MCP: The LLM Translates, CDS ExecutesThe division of labour is the argument, and it is in the subtitle: the model turns English into a CDS query and CDS runs it, so nothing generates SQL and nothing hallucinates a join. Schema annotations do the resolving. Long, with the package, the repository and the MCP registry entry all linked.Shahid, published 2026-06-25
Banking Sentinel: A Multi-Agent Credit-Risk Copilot on SAP BTP, Built on a Mixed AI StackBanking Sentinel: A Multi-Agent Credit-Risk Copilot on SAP BTP, Built on a Mixed AI StackThe risk it is chasing is a graph problem stated in banking terms: a borrower who looks safe alone and is the linchpin of a group ten times the size. Multi-agent, mixed AI stack, on BTP and CAP.Shahid, published 2026-06-24
Banking Sentinel: A Multi-Agent Credit-Risk Copilot on SAP BTP, Built on a Mixed AI Stack
Agent tooling changed in @cap-js/agents 0.9.1, 2026-08-14. Affects A recipe, Runnable code.
Beyond Basic AI Error Handling in SAP CPI - A Rules-First Architecture with HANA Cloud LearningBeyond Basic AI Error Handling in SAP CPI - A Rules-First Architecture with HANA Cloud LearningBuilds a rules-first error-diagnosis pipeline on Cloud Integration in SAP Integration Suite: a Groovy classifier resolves known failures — auth, timeout, malformed payload — deterministically, and only genuinely novel errors reach an LLM through the AI Adapter, after a sanitizer script replaces every payload leaf value with a typed placeholder before it leaves the tenant. SAP HANA Cloud, reached over JDBC, holds a fingerprinted dedup cache and an accumulating table of past diagnoses. It also documents the AI Adapter configuration — Advanced mode, a Bearer-prefixed Secure Parameter, and a connection/processing URL split — needed to reach a non-SAP OpenAI-compatible endpoint, found by the author to be undocumented.Krishna_Karthik, published 2026-06-11
Building an LLM Agent in CAP Java with Spring AIBuilding an LLM Agent in CAP Java with Spring AIThe Java counterpart to the author's own Node.js post, and written as one: connecting to AI Core with the SAP AI SDK, then calling an agent and a tool through Spring AI. Opens with a section on libraries and workflows — which framework does what, and what is left to you — before any code.published 2026-06-08
Building an LLM Agent in CAP Java with Spring AI
Agent tooling changed in @cap-js/agents 0.9.1, 2026-08-14. Affects Runnable code.
Calling Claude from ABAP: How I Wired SAP AI Core into a Real Business ScenarioCalling Claude from ABAP: How I Wired SAP AI Core into a Real Business ScenarioA proof of concept wiring SAP AI Core's Claude deployment directly into ABAP with no CAP layer between them: two SM59 HTTP destinations, a hand-written OAuth 2.0 client-credentials exchange, and a regex to unwrap Claude's system/messages response envelope. Most of the effort goes into a system prompt that constrains the model to return one fixed JSON schema of filter fields, which ABAP deserialises with /UI2/CL_JSON and exposes through a RAP custom action so a Fiori search field can drive an OData V4 filter.PranayGaroo, published 2026-06-04
AI-native North Star architecture — Platform LayerAI-native North Star architecture — Platform LayerThe platform layer of SAP's North Star architecture, the chapter that says where agents run and what a developer builds them with. It names an AI Golden Path built on the SAP Cloud Application Programming Model, the ABAP Cloud development model with RAP and the SAP Fiori design system, and a managed agent runtime on SAP BTP that is designed to provide software development kits for building agents in any framework, an agent sandbox, container-hosted execution, a skills registry and continuous evaluations, which is the ground agent tooling covers. Agents are extended without changing their code through four patterns, the first of them tool extensibility through the Model Context Protocol, the concern MCP servers is named for; the process-layer chapter adds that a pro-code path with a standardised agent SDK and a prompt-driven path through the Joule Studio solution produce the same code artifact on the same runtime. It describes a design SAP intends rather than products that exist, and the SDK and the registry it names have no entry here yet.published 2026-05-13
AI-native North Star architecture — Platform Layer
Agent tooling changed in @cap-js/agents 0.9.1, 2026-08-14. Affects A reference.
Agentic AI on BTP: a Multi‑Agent App with Pydantic AIAgentic AI on BTP: a Multi‑Agent App with Pydantic AISplits a single CAP/TypeScript agent that had grown to juggle five BTP MCP servers behind one bloated tool list into btp-multiagent-app, a Python and Pydantic AI rewrite where the Cloud Foundry and BTP Core MCP servers each get a dedicated specialist agent, with an audit log specialist scaffolded but left disabled, behind an orchestrator whose only tools are delegation calls to those specialists, which is the agent-tooling pattern the piece works through. Each specialist authenticates to its MCP server over Streamable HTTP with OAuth 2.1, and a shared module centralises the SAP AI Core GenAI Hub model setup behind two credential layouts, one for local runs and one for the Cloud Foundry service binding, the LLM integration side of the rewrite. Tool lists per agent drop from well over a hundred to a dozen or two, and routing failures become separable from tool-misuse failures by reading the trace.published 2026-05-03
Agentic AI on BTP: a Single Agent with LangGraphAgentic AI on BTP: a Single Agent with LangGraphA CAP TypeScript agent puts a LangGraph StateGraph in front of AzureOpenAiChatClient from @sap-ai-sdk/langchain, reaching SAP AI Core's Generative AI Hub, and lets its tool-calling loop draw on five deployed MCP servers — OData MCP Proxy, CI MCP Server, AI Core MCP Server, BTP MCP Server and BTP CF MCP Server — through @langchain/mcp-adapters' MultiServerMCPClient. Because agent, servers and APIs share one XSUAA tenant, the chat service forwards the signed-in user's own JWT instead of a service account, so each downstream server enforces that user's own authorizations. Past roughly a hundred combined tools the single prompt slows down and stops separating concerns, which is why the author's next post splits it into one specialist agent per MCP server behind an orchestrator.published 2026-05-03
Enhancing Enterprise Applications: AI-powered User Experience with SAP CAP AI PluginEnhancing Enterprise Applications: AI-powered User Experience with SAP CAP AI PluginStarts from the blank form nobody wants to fill in, and asks what an enterprise application should do about it. The AI is in service of that rather than the subject of it.published 2026-04-29
What Happens to Your Data When You Enable SAP's AI Features?What Happens to Your Data When You Enable SAP's AI Features?Traces where a prompt goes when Joule, the Generative AI Hub or the orchestration service calls a third-party model through SAP AI Core: the prompt and the business context around it travel to a model such as GPT-4o, Claude or Gemini for inference, and masking through SAP Data Privacy Integration is an optional orchestration module rather than a default, so with no masking block in the @sap-ai-sdk/orchestration client config nothing is masked. On the masking options, which the author says he confirmed with SAP AI Core product management, anonymization discards the original value and pseudonymization keeps it reversible; DPI recognizes person names, email addresses and phone numbers, plus custom entities defined by regex, which leaves commercially sensitive non-PII such as a revenue forecast or a project codename to pass through unless someone writes a pattern for it, and whether that data is covered at all is one of three answers the post marks as still pending. The "no training" assurance for third-party providers is read as contractual rather than technical, and models hosted and managed by SAP, Mistral among them, are named as a path other than a US-hyperscaler-hosted one.jaripie, published 2026-04-23
Custom Joule Agent for SAP Ariba SLP - Building an AI Agent for Supplier Registration ApprovalCustom Joule Agent for SAP Ariba SLP - Building an AI Agent for Supplier Registration ApprovalBuilds a Joule Studio agent that reviews SAP Ariba SLP supplier registration tasks: four custom skills bind to Ariba's API (pending tasks, task detail, questionnaire answers, post a decision), each needing a precisely configured BTP destination and action, and the agent's own Expertise, Instructions and Additional Context sections carry the domain knowledge, the step order, and the scoring rubrics that decide approve, deny or deny-with-resubmit. The author contrasts this with an earlier hand-wired flow on SAP Build Process Automation, where every connection, variable and error path had to be designed and wired manually.Giancarlo-Martinez, published 2026-04-07
SAP Agentic AI in Practice: Building the SAP Exception Log Triage Agent (Part 3 of 6)SAP Agentic AI in Practice: Building the SAP Exception Log Triage Agent (Part 3 of 6)Part 3 of a six-part series builds a router agent that classifies incoming SM21 system log errors and SM37 job failures, delegates each to a specialized sub-agent built with LangGraph's create_react_agent, and correlates findings across both logs into a single severity score and escalation list. State moves through a typed dict shared between the router and its sub-agents, with tools that query SAP HANA tables directly; a reactive on-demand mode sits alongside a scheduled monitoring loop that compares error counts to a baseline and escalates automatically past a threshold. The router-plus-sub-agent split carries forward to Part 6, where the same router coordinates the assessment agents built in the series' other parts.DEEPA_DORAIRAJ, published 2026-04-06
Integrating a Custom AI Model into SAP CAP Using AI LaunchpadIntegrating a Custom AI Model into SAP CAP Using AI LaunchpadYour own model rather than a hosted one: deployed to AI Core, made callable through AI Launchpad, then reached from CAP. The case a lot of companies ask about and few posts cover.egeAksoyek, published 2026-03-20
Integrating a Custom AI Model into SAP CAP Using AI Launchpad
LLM integration changed in @cap-js/ai 1.0.0, 2026-04-28. Affects Runnable code.
Integrating and Extending JouleIntegrating and Extending JouleSAP's reference architecture for SAP Joule across SAP cloud applications: a request is checked against a Scenario Catalog of the applications' skills and a Knowledge Catalog answered by retrieval-augmented generation, combined with the user's roles and history, sent to an LLM available through the generative AI hub, and, where a Joule function is needed, turned into a call to the backend. Its identity page is the authentication path: each user sits in the Identity Directory of Cloud Identity Services with a Global User ID, Identity Provisioning synchronises role assignments into SAP Build Work Zone, whose navigation service decides which functions a user may call, and the Destination service exchanges the user's token to call a function in the user's name. Its extension page covers Joule Studio in SAP Build, where Joule skills automate rule-based tasks over APIs and AI agents plan and reason with models and document grounding from SAP AI Core, the ground agent tooling covers; a later Joule Studio page describes a managed Joule Studio Runtime on the SAP Business AI Platform, which has no entry here.published 2026-03-16
SAP Integration Suite: Intelligent Integration with AI AdapterSAP Integration Suite: Intelligent Integration with AI AdapterThe AI Adapter adds generative AI as a receiver step inside a Cloud Integration flow, connecting through either AICore-GenAIHub — a single API onto SAP AI Core's Generative AI Hub and its foundation models — or Custom-OpenAI-Rest, a plain REST call out to a provider such as OpenAI. SAP Integration Suite starts the call from the iFlow and feeds the model's response into the next step, which is the mechanics behind the sentiment-analysis, summarisation and routing use cases the piece lists without walking through any one of them.meena_choudhary, published 2026-01-28
cds-toon-ai-protocol-adapter: a pragmatic way to cut AI token costs using SAP CAPcds-toon-ai-protocol-adapter: a pragmatic way to cut AI token costs using SAP CAPNames a cost most CAP-plus-LLM writing ignores: every JSON payload streamed to a model is billed by its verbosity, in latency as well as money. The plugin brings TOON — Token-Oriented Object Notation — to a CAP service so the wire format changes without the service code doing so, in a REST variant and an OData one, with token counts reported on both sides.published 2026-01-25
UI5 MCP Server Got Extended for UI Integration Cards GenerationUI5 MCP Server Got Extended for UI Integration Cards GenerationVersion 0.2.1 of the UI5 MCP Server adds three tools for UI Integration Cards: get_integration_cards_guidelines feeds an agent current best practices and API versions, create_integration_card scaffolds a new card project, and run_manifest_validation checks the generated manifest against the schema. The post frames these as closing the same gaps as the server's existing tools — an LLM's outdated or wrong card API knowledge — this time for scaffolding and validating a card rather than a full app.daniel_vladinov, published 2026-01-22
SAP JOULE & ME : My First JOULE Skill in SAP EHSM-Incident Management Without CodingSAP JOULE & ME : My First JOULE Skill in SAP EHSM-Incident Management Without CodingWalks through building a transactional Joule Skill in Joule Studio, classic edition without writing code: activating an EHS Incident OData service in SAP S/4HANA, wrapping it as an Action Project with chosen input and output fields, then creating, saving and deploying the skill from those fields. Tested against a live system, the deployed skill lets a user create an EHS incident and list reported incidents through Joule's conversational interface or a configured button.Pardhasaradhi_Reddy_Chelikam, published 2026-01-05
Enhancing Business Processes with SAP BTP AI in SAP Integration SuiteEnhancing Business Processes with SAP BTP AI in SAP Integration SuiteWalks through a post-exit iflow in SAP Integration Suite that calls an LLM through SAP AI Core's Generative AI Hub to turn a customer's free-text return comment into one of a fixed set of SAP rejection-reason codes, so an inbound ORDERS05 IDoc carries a decision (E1EDP01-ABGRU) the ERP can act on without pausing for manual review. The worked example is a Commerce Cloud returns flow into S/4HANA; the piece argues the same bounded-classification pattern fits other narrative fields — delivery-exception remarks, expense-claim justifications — where the cost of an LLM call is repaid by removing a human review step, and points to a colleague's post on the Generative AI Hub orchestration workflow for the guardrailing (prompt templating, data masking) it only names.arunmaarirajha_kv, published 2025-12-15
Prompt-Driven Development: Building Full-Stack SAP CAP Apps from User StoriesPrompt-Driven Development: Building Full-Stack SAP CAP Apps from User StoriesA user story in, a scaffolded full-stack CAP application out, through SAP Build Code and Joule. The question it is actually asking is how much of the foundational code is worth generating, and it answers by generating all of it.AnuradhaDhingan, published 2025-12-11
Event Driven SAP CAP on Kyma with Agentic AI and UI Auto RefreshEvent Driven SAP CAP on Kyma with Agentic AI and UI Auto RefreshA prototype, described as one: a plain CAP application made to react to external events on Kyma, with an agent in the loop and the UI refreshing itself rather than being reloaded. Short — it reports what was built and why, not how to rebuild it.Shahid, published 2025-12-01
Streaming LLM Output with CAP and WebSocketsStreaming LLM Output with CAP and WebSocketsTokens arriving from the AI SDK's orchestration module, forwarded to the browser as they come rather than after the model finishes. Covers all four pieces — enabling WebSocket support in CAP, a service with actions and events, the streaming itself, and a frontend on the native WebSocket API — which is what makes it the clearest case for why a CAP service would want a socket at all.RobinP, published 2025-11-18
Streaming LLM Output with CAP and WebSockets
LLM integration changed in @cap-js/ai 1.0.0, 2026-04-28. Affects A recipe, Runnable code.
How to Reduce Prompt Token Costs Using Toon = Save MoneyHow to Reduce Prompt Token Costs Using Toon = Save MoneyToken spend treated as a running cost rather than a detail, and TOON as the lever. Companion reading to the CAP adapter for the same format that is indexed here.Yogananda, published 2025-11-12
How to Reduce Prompt Token Costs Using Toon = Save Money
LLM integration changed in @cap-js/ai 1.0.0, 2026-04-28. Affects A recipe, Runnable code.
How Redwood Software and SAP are shaping the next era of automation with Joule StudioHow Redwood Software and SAP are shaping the next era of automation with Joule StudioRedwood's Field CTO describes a tech preview where an SAP Basis administrator tells SAP Joule in natural language to pause background jobs and dependent Integration Suite iFlows for unplanned system maintenance; Joule acts as a control tower, triggering a pre-approved RunMyJobs workflow and asking for approval before the restart proceeds. The integration is built by publishing RunMyJobs operations to the SAP Business Accelerator Hub, securing access through SAP BTP with API keys and role-based permissions, and combining those actions into a skill inside Joule Studio's low-code design canvas.gmblom, published 2025-11-02
Generative AI on SAP BTPGenerative AI on SAP BTPThe reference shape for a generative AI application on BTP: a CAP backend, SAP HANA Cloud holding data and embeddings for similarity search through its vector engine, and the Generative AI Hub in SAP AI Core as the single access point to foundation models. It shows retrieval-augmented generation on that stack and names the CAP plugins and SDKs that connect the pieces, for both Cloud Foundry and Kyma.published 2025-09-02
Generative AI on SAP BTP
LLM integration changed in @cap-js/ai 1.0.0, 2026-04-28. Affects A reference.
How to use Cloud SDK for AI in an existing CAP NodeJS Project in CommonJSHow to use Cloud SDK for AI in an existing CAP NodeJS Project in CommonJSA dependency that is ESM-only inside a project that is CommonJS, and the author's judgement that converting the project was too risky. What follows is the esbuild workaround, which is the honest answer when the ecosystem has moved and your codebase has not.Jacky_Liu1, published 2025-08-01
Improving Exception Handling using Generative AI in SAP Cloud Integration (SAP CI/CPI)Improving Exception Handling using Generative AI in SAP Cloud Integration (SAP CI/CPI)Wires an Exception Subprocess in Cloud Integration to send a failed iFlow's caught exception text to Google Gemini's generateContent endpoint over the HTTP adapter, then parses the response in Groovy for a user-friendly error message, the likely causes, the steps to try and an HTTP status code to return to the sender. A Content Modifier holds the API key and the prompt as externalised properties, which the post says is what lets another model be put in their place, and a closing Groovy script writes the extracted fields to custom headers and log attachments and rebuilds them as an XML error response. The exception-subprocess wiring is given step by step and so is the Gemini call with its request format, which is why the piece sits on integration flows and on LLM integration.AdarshRao, published 2025-07-20
Lets Talk to the Database - CAP & SQLite Powered by Google Gemini-2.5-ProLets Talk to the Database - CAP & SQLite Powered by Google Gemini-2.5-ProNatural language against a CAP service on SQLite, with a non-SAP model doing the translation. Written as a learning exercise rather than a pattern, and the choice of Gemini over anything in the AI hub is the interesting part of it.ShivamShuklaSAP, published 2025-04-16
Elevating Data Privacy in SAP CAP based GenAI Applications with CAP LLM PluginElevating Data Privacy in SAP CAP based GenAI Applications with CAP LLM PluginThe question underneath the enthusiasm: what leaves your landscape when a prompt goes to a model. The plugin's answer is to anonymise the sensitive fields through HANA before the prompt is sent, and this is the post that explains why that is the design rather than an option.AkashAmarendra, published 2024-02-02
An Innovative Language Exchange App developed with SAP Build Apps, CAP, Kyma, and OpenAIAn Innovative Language Exchange App developed with SAP Build Apps, CAP, Kyma, and OpenAIA dormant internal app at SAP Japan revived with AI, built to solve a problem the team actually had — colleagues who could not talk to each other. Build Apps on the front, CAP on Kyma behind it, OpenAI in the middle.javieryw, published 2024-01-31
An Innovative Language Exchange App developed with SAP Build Apps, CAP, Kyma, and OpenAI
LLM integration changed in @cap-js/ai 1.0.0, 2026-04-28. Affects Runnable code.
GenAI Mail Insights - Leveraging the generative AI hub in SAP AI Core to improve customer supportGenAI Mail Insights - Leveraging the generative AI hub in SAP AI Core to improve customer supportA sample application rather than an explainer, and the author says so in the opening — enough talking, here is something that runs. Incoming support mail summarised and categorised through the generative AI hub in AI Core.martinfrick, published 2023-11-15
Automate your Job Postings Using Generative AI (ChatGPT) & SAP Build Process Automation AppsAutomate your Job Postings Using Generative AI (ChatGPT) & SAP Build Process Automation AppsA process that drafts its own text: ChatGPT writing job postings, SAP Build Process Automation carrying them through approval, and CAP in between. Presented as a generic pattern the author expects to be adapted rather than copied.shivamshukla12, published 2023-08-10
Creating a meal generator application with SAP Build, Cloud Application Programming and openAICreating a meal generator application with SAP Build, Cloud Application Programming and openAIIngredients in, meal suggestions out, assembled from BTP services plus OpenAI and Google APIs. A small application built for the pleasure of it, and useful here as the plainest example of CAP holding the middle of a generative flow.former_member793421, published 2023-07-17
Empowering Your Data Analytics with GPT-3 and SAP CAP IntegrationEmpowering Your Data Analytics with GPT-3 and SAP CAP IntegrationFree text in, SQL out, answer back in a sentence — the model writes the query, CAP runs it, and the model turns the rows into prose. Written in the GPT-3 window, so the model access has dated, but the three-step shape has not, and it is a different bet from retrieval: nothing is embedded, the schema is the context.Former Member, published 2023-03-10
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 3/3A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 3/3The SAPUI5 frontend of the sample, and mostly an argument for writing UI5 in TypeScript: ES imports in place of sap.ui.define, a class with real private members in place of the underscore convention, and the service types generated from the CDS model copied into the UI package, so one IChats interface types the OData object the controller reads. The second half is structure — a formatter class and a singleton LayoutManager called from the XML view through core:require, a confirmation helper taking a callback, a dialog class resolving a promise, a generic createEntity<T> on a service class — all of it to keep every controller under 100 lines. What it says about arranging a UI5 app holds whatever model sits behind the chat; the tooling around it is of its date, and the post calls UI5's TypeScript support experimental.published 2023-02-27
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 3/3
cds2types. Affects Runnable code.
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 2/3A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 2/3The backend: CAP with TypeScript, the entities holding chats and personalities, and the service that talks to the API. This is where the streamed completion and the repositories behind the handler come from, which is the part of the sample still worth lifting even though the model access has since been superseded by AI Core.published 2023-02-24
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 2/3
cap-ui5-gpt-chat: no published releases, repository quiet for 41 months. Affects Runnable code.
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 1/3A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 1/3The concepts before the code: what the GPT model and its API actually offer, and what the finished chat application looks like — personalities included. Written in the weeks when the ground was moving under it, and the post carries its own change log to prove it: GPT-3.5 added on 1 March, GPT-4 on 15 March, and an admission that a series planned as two posts became three.published 2023-02-22
A ChatGPT-like chat app built with modern SAP technologies (CAP, SAPUI5) – Part 1/3
cap-ui5-gpt-chat: no published releases, repository quiet for 41 months. Affects Runnable code.
Content-based Agents with SAP Joule Studio Agent BuilderContent-based Agents with SAP Joule Studio Agent BuilderJoule Studio, classic edition, the capability in SAP Build the page calls Joule Studio, is where an organisation creates and manages its own Joule skills and Joule agents; its Agent Builder is for a custom agent that handles a multi-step task by combining instructions, model-backed planning and reasoning, and a configured set of tools, Joule skills, document resources, MCP servers, or other agents run as subagents. The page states an agent executes inside the calling user's own security context through SAP Cloud Identity Services, and can be tried against a test conversation before deployment. That is the tool-and-turn-taking ground agent tooling covers, here built through Joule Studio rather than in application code.
Pro-Code AI AgentsPro-Code AI AgentsPro-code agents are the most flexible way the page frames building an agentic workflow on SAP BTP: full autonomy, custom API integrations and algorithms chosen by the developer, set against declarative alternatives such as content-based agents and Joule Scenarios. It names LangGraph as the framework it recommends for production, enterprise-grade agentic workflows, alongside the ReAct pattern as the plan-act-observe loop such an agent runs. That is the ground agent tooling covers, building an agent's own tool-calling loop rather than exposing an application to somebody else's agent.
Generate agentic workflows using SAP-managed n8n in Joule StudioGenerate agentic workflows using SAP-managed n8n in Joule StudioSAP-managed n8n inside Joule Studio generates a workflow from a business intent a developer states, and the workflow is then tested and deployed, consuming line-of-business application APIs, events and agents in SAP Business AI Platform: process automation with the workflow generated from intent, and agent tooling where the workflow calls agents. The generated workflow sends notifications to the Joule Work mobile app of SAP Joule, to Microsoft Teams or to Slack. SAP's abstract names the pieces and not how the generation or the n8n hosting works. A two-hour hands-on TechEd 2026 session (Tuesday 27 October, Berlin), presented by Rama Sulakshani Murari, Rumen Dimitrov, Seshadri Sreenivas Ramanarayanan and Venugopal Chembrakalathil, written here from SAP's abstract; the recording is not yet published.
Connecting AI agents to the ABAP RESTful application programming modelConnecting AI agents to the ABAP RESTful application programming modelJoule Studio, classic edition connects Joule Agents to applications built with the ABAP RESTful application programming model through actions and skills, calling the OData services those applications expose their business objects through: the tools an agent may call, which is the agent tooling concern seen from the ABAP side. The abstract promises alternative integration patterns and does not name them. A TechEd 2026 session (Tuesday 27 October, Berlin), presented by Leon Karlau, written here from SAP's abstract; the recording is not yet published.
Building agentic applications with SAP HANA CloudBuilding agentic applications with SAP HANA CloudA CAP application extended with semantic search and agent features on top of SAP HANA Cloud's vector engine and SAP AI Core, built in a hands-on workshop. The abstract names the parts, CAP, the vector engine and AI Core, and the two additions: semantic search, the retrieval half of retrieval-augmented generation, and LLM-supported agent capabilities inside the application, which is agent tooling. Which embedding model, which CAP runtime and how the agent calls its tools are not stated. A TechEd 2026 session (Thursday 29 October, Berlin), presented by Hui Li, Ritty Joseph, Shraddha Shetty and Tanuja Jadhav, written here from SAP's abstract; the recording is not yet published.
Agentic memory: Knowledge graph engine and recall in SAP HANA CloudAgentic memory: Knowledge graph engine and recall in SAP HANA CloudThe knowledge graph engine in SAP HANA Cloud as an agent's memory: what the engine, the graph side of graph, spatial and document data, does in an agentic workflow, how an agent keeps what it learns there, the state agent tooling holds between turns, and the SAP AI Core integration through which the database calls LLMs and tabular foundation models such as SAP-RPT-1. The abstract names the engine, the memory role and the AI Core link and does not say how the memory is written or read. DA263 on Wednesday 28 October is a two-hour hands-on on the same engine with a different cast. A TechEd 2026 session (Tuesday 27 October, Berlin), presented by Markus Fath and Shraddha Shetty, written here from SAP's abstract; the recording is not yet published.
Knowledge graph and agent memory: Build an agent that remembersKnowledge graph and agent memory: Build an agent that remembersThe knowledge graph engine in SAP HANA Cloud as an agent's memory: what the engine, the graph side of graph, spatial and document data, does in an agentic workflow, how an agent keeps what it learns there, the state agent tooling holds between turns, and the SAP AI Core integration through which the database calls LLMs and tabular foundation models such as SAP-RPT-1. The abstract names the engine, the memory role and the AI Core link and does not say how the memory is written or read. DA205 on Tuesday 27 October is the 45-minute talk on the same engine, with a near-identical abstract and a different cast. A two-hour hands-on TechEd 2026 session (Wednesday 28 October, Berlin), presented by Christoph Morgen, Jan Zwickel, Markus Fath and Thomas Hammer, written here from SAP's abstract; the recording is not yet published.
Access to Generative AI ModelsAccess to Generative AI ModelsSets out the plain way to call a deployed generative AI model on SAP BTP: a request that carries the prompt, the model and the parameters that control generation, with the same code able to switch between models. It lists data masking, prompt templating and filtering as features that come with the access rather than as code the caller writes, which is the shape LLM integration takes on the BTP door; the page's implementation sections sit behind its steps and were not read.
Agentic CAP: Build MCP servers and enterprise AI agentsAgentic CAP: Build MCP servers and enterprise AI agentsCAP applications become AI agents over the Model Context Protocol and the Agent2Agent protocol, which is the serve your application over MCP concern, and SAP AI Core is the LLM integration through which AI is embedded in a CAP app, with recommendations from the SAP-RPT-1 model as the abstract's example. The abstract names the two protocols and the one model and does not say which packages carry them. A two-hour TechEd 2026 session in a hands-on room (Tuesday 27 October, Berlin), presented by Daniel Hutzel, Daniel Schlachter, Marc Huber and Sami Lechner, written here from SAP's abstract; the recording is not yet published.
AI-enabled administration: Automating log analysis with generative AI hubAI-enabled administration: Automating log analysis with generative AI hubA log analysis tool for incident work, built in SAP Business Application Studio with the generative AI hub feature, which SAP AI Core ships as its LLM integration by its own documentation: it automates root-cause analysis over system errors and masks sensitive data. SAP's abstract states the tool's jobs and not how the hub is called, what the masking covers or which model answers. A TechEd 2026 session (Wednesday 28 October, Berlin), with no speaker listed in the catalogue at the time of writing, written here from SAP's abstract; the recording is not yet published.
Stop guessing, start counting: How generative AI fixed an inventory messStop guessing, start counting: How generative AI fixed an inventory messAn inventory-counting application a business built on SAP Business AI Platform with SAP HANA Cloud and generative AI, to replace recount lists kept by gut feel across SAP ERP Central Component, SAP S/4HANA and Microsoft Excel; the talk promises the app, its architecture and what went wrong. The generative AI part is LLM integration from application code; the abstract names HANA Cloud and generative AI as the two building blocks and says nothing about how either is used. A TechEd 2026 session (Thursday 29 October, Berlin), presented by Ranjeet Kumar, written here from SAP's abstract; the recording is not yet published.
CAP in the age of AI #5 - agents wrap-upCAP in the age of AI #5 - agents wrap-upThe agents topic closed out on what sits around the agent rather than inside it, now that the documentation has moved into capire. The chat preview's session is persisted in the database rather than kept in memory, traces go out through @cap-js/telemetry to a self-hosted MLflow under the OpenTelemetry conventions for generative AI, and the panel is explicit that @agent reuses the MCP adapter's code programmatically without serving an MCP endpoint. It ends in the xtravels sample, where a travel agent delegates to a hotel-service sub-agent and reaches a separately started flight service over MCP through the service binding CAP resolves anyway. Earlier in the hour the same agent answers from a model running on the presenter's laptop, reached as an OpenAI-compatible endpoint.published 2026-10-02
CAP in the age of AI #4 - more on agentsCAP in the age of AI #4 - more on agentsAsking a human first, and telling the agent how to behave. Ordering a book through @cap-js/agents happens without a question being asked; the human-in-the-loop annotation on the action turns the same request into an approve-or-reject step, which the episode follows down to the A2A status update that stores the task in an input-required state in the database — A2A itself being JSON-RPC 2.0 over HTTP with server-sent events for the stream. The rest is instructions on disk: an agents.md in the service's own folder removes the question about how many copies to order, a skill under a skills folder is advertised by its description and read only when the agent decides the case applies, and either file makes the plugin build a deep agent instead of the plain one.published 2026-09-11
CAP in the age of AI #3 - from MCP to A2ACAP in the age of AI #3 - from MCP to A2AThe episode that finishes @cap-js/mcp with authorisation and then crosses to agents. @requires and @restrict are the same annotations CAP already has, and the plugin reads them to filter the tool list and the describe output per user and per request, so a model is not told about entities the caller cannot reach — where the OData metadata of the same service lists every exposed entity regardless. The second half annotates a service with @agent from @cap-js/agents, which serves it over A2A with an agent card and a chat preview, runs first against a mocked response and then against SAP AI Core bound with cds bind and a hybrid profile, and reuses the MCP adapter's tools rather than defining new ones.published 2026-08-21
CAP in the age of AICAP in the age of AISeven live episodes of Hands-on SAP Dev in which DJ Adams works through CAP's AI features with the people who build them, each one a screen-share against a sample project. The first three are @cap-js/mcp: the @mcp annotation on a service, the describe, query and call tools, CQL as the query format, and what authorisation takes out of the tool list. From episode three the run moves to A2A and @cap-js/agents — human approval of an action, agents.md and skills, session persistence, telemetry and sub-agents — and the last two episodes are vector embeddings in CDS, local on SQLite and native on SAP HANA Cloud.published 2026-08-07
Building an LLM Agent in CAP Java with Spring AIBuilding an LLM Agent in CAP Java with Spring AIThe Java counterpart to the author's own Node.js post, and written as one: connecting to AI Core with the SAP AI SDK, then calling an agent and a tool through Spring AI. Opens with a section on libraries and workflows — which framework does what, and what is left to you — before any code.published 2026-06-08
Building an LLM Agent in CAP Java with Spring AI
Agent tooling changed in @cap-js/agents 0.9.1, 2026-08-14. Affects Runnable code.
CAP in the age of AI #7 - vector embeddings wrap-upCAP in the age of AI #7 - vector embeddings wrap-upThe same embeddings model taken from SQLite to SAP HANA Cloud, then handed to an agent. Locally it is @cap-js/ai that adds the vector functions SQLite lacks and downloads an all-MiniLM L6v2 ONNX model on first start; on HANA natural language processing has to be switched on, the model name is a mandatory argument, the vectors it returns are 768 dimensions, and dimensions must match on both sides of a comparison. HANA itself can be given SAP AI Core as a remote source for other embedding models, which a CDS.ql expression has to inject because the postfix notation only takes what the CDS model holds, and the hour ends with a search function over cosine similarity exposed over @cap-js/mcp, so an agent asked about low output from solar panels calls semantic search instead of writing the query itself.published 2026-10-16
CAP in the age of AI #6 - vector embeddingsCAP in the age of AI #6 - vector embeddingsVector embeddings put to work in the incident management sample, answering whether this problem has been seen before. An embedding element of type Vector is calculated and stored from the vector_embedding function, so the database refreshes it whenever the text behind it changes, and cosine_similarity compares it with a freshly embedded string; because a title says little, a scheduled action first summarises each incident's conversation through a model on SAP AI Core and the embedding is taken from that summary. The similar-incidents list is a view joining incidents to themselves, with order and limit set in a before handler on the expanded query, and the vector element excluded from the service, since OData knows nothing about vectors — on SAP HANA Cloud the function is native, while SQLite needs packages added to the project.published 2026-10-09
CAP in the age of AI #2 - more on the MCP adapterCAP in the age of AI #2 - more on the MCP adapterWhat the query tool costs in tokens, and the third tool @cap-js/mcp registers. Counted against a model running locally under llama.cpp, the same query is 26 tokens as CQL and 173 as CQN, and the tool definition sent with every system prompt drops from around 5,500 tokens to around 600 — the reason CQL became the default format. The call tool then exposes a service's unbound actions and functions, as one generic tool or, behind a flag, one tool per action with its own description and input schema; a question from the chat about extension points gets the answer that the adapter does the tools, which the panel calls a small subset of the protocol.published 2026-08-14
CAP in the age of AI #1 - the MCP adapterCAP in the age of AI #1 - the MCP adapterThe episode that opens the series and the first pass over @cap-js/mcp, with four of the people building it on the call. Two of the adapter's tools are worked through: describe, the one an MCP client reaches for first to get the overview of what it may do, and query, which the panel establishes as read-only. It also writes the server entry into a local agent's own configuration file, which episode three comes back to as auto-wiring.published 2026-08-07
Details
Graph
Agent toolingcapability, core and community
LLM integrationcapability, core and community
SAP Communitypublishing source
SAP Communitycap-tag
Source
Type
blog
Published on
Host
community.sap.com
Discovered via
First seen
History
- explains → applies_to on capability:agent-tooling (cap, btp); capability:llm-integration (cap, btp)Door scope proposed from the title and summary, so a door page lists the piece only where it holds (docs/concepts/content-on-doors.md).
- Entered the atlasDiscovered via SAP Community · cap-tag.
For your agent
- Ref
content:building-an-llm-agent-using-cap- JSON
- /content/building-an-llm-agent-using-cap.json
- MCP
- Connecting an agent