Dataiku vs OutSystems (2026)

Agents call managed tools, API endpoints, OpenAPI specs, local and remote MCP servers, Knowledge Bank search, LLM Mesh queries, web search, inline code and plugin tools, over connections secured per group.

Sourcedoc.dataiku.com/dss/latest/agents/index.htmlread 2026-09-27

Agents call tools directly. Server actions are exposed to an agent, and the AI model autonomously decides which action or sequence of actions is appropriate by analyzing the prompts and the purpose of each exposed action, calling a single action, a sequence, or the same action repeatedly.

Agents connect to APIs, databases, ERP, CRM and legacy systems through prebuilt connectors and extensible services, wrapping and extending core systems without replacement, and Model Context Protocol support gives agents direct access to enterprise systems, external tools and services. Forge components extend the tool surface, such as OmniDoc2MD, whose server action can be exposed as an agent action.

Sourcesuccess.outsystems.com AI agent actions documentation, outsystems.com Agent Workbench GA announcementread 2026-08-31

Structured visual agents define sequences with routing, delegation to other agents, parallel and for each branches, agentic loops and reflection blocks, alongside code agents and external agents from Snowflake Cortex, Databricks, AWS Bedrock and Google Vertex AI; Agent Hub can orchestrate a query across several agents.

Sourcedoc.dataiku.com/dss/latest/agents/structured-visual-agents/index.htmlread 2026-09-27

Named multi agent architectures include a shared memory pattern, where agents read from and write to shared state and communicate implicitly by modifying it, and a network pattern of autonomous peer agents interacting in a graph like structure without a central controller.

Agents combine with ODC workflows, where automatic activities call an agent's service action, human activities pause for approval, and validation branches route invalid results back to a prior step to rerun the agent with feedback. Agents are coordinated side by side, in sequence, in hierarchy, and with or without humans in the loop, through a visual interface. Agentic Enterprise Orchestration, launched at ONE, is the current generation of this layer.

Sourcesuccess.outsystems.com agentic patterns and building agentic workflows documentationread 2026-08-31

Agents run inside automation scenarios, which start on time based, dataset change, SQL query change, scenario completion or Python triggers without a person, and users reach agents in Slack (DM or mention), Microsoft Teams, Agent Hub, Agent Chat and through APIs.

Sourcedoc.dataiku.com/dss/latest/scenarios/index.htmlread 2026-09-27

Agent runs start from three entry routes, an Event Handler, a Timer or a REST Expose. Timers are background scheduled jobs configured with OutSystems schedule syntax and forced on demand through the Wake action, events run through the platform's event driven architecture, and REST Expose makes an agent callable as an inbound HTTP endpoint. Agents are additionally published as service actions consumed by any app in the tenant, embedded into web and mobile interfaces, and reachable externally through the Agent Experience MCP and A2A services.

Sourcesuccess.outsystems.com monitoring and troubleshooting apps, use timers and event-driven architecture documentationread 2026-08-31

Knowledge Banks embed datasets and documents into vector stores that agents search through a Knowledge Bank Search tool with static, dynamic and agent inferred filters and document level security, with GraphRAG and automated RAG optimization also documented.

Sourcedoc.dataiku.com/dss/latest/agents/tools/knowledge-bank-search.htmlread 2026-09-27

Agents are grounded in the OutSystems Enterprise Context Graph, an operational knowledge layer over the customer's application estate that OutSystems says grounds any coding tool or LLM in shared context, spanning entities, relationships, roles and business logic.

Agent configuration begins with grounding data that connects the agent to internal databases or external APIs so responses rest on current factual information, and integrated semantic search shipped with Agentic Enterprise Orchestration at ONE.

Forge components ingest documents, converting Word, PDF, spreadsheet, presentation, HTML and plain text binaries to Markdown while preserving structure and extracting images.

Sourcesuccess.outsystems.com creating an agent and AI agent actions documentation, outsystems.com ONE 2026 platform announcementread 2026-08-31

Long term memory blocks store facts, such as preferences, and episodes from past conversations in memory banks, kept separately per user, shared by agents in a project through a default bank or split into named banks, and retrieved by LLM filtering and semantic search in later conversations. No retention period or delete path is stated.

Sourcedoc.dataiku.com/dss/latest/agents/structured-visual-agents/blocks/long-term-memory.htmlread 2026-09-27

OutSystems offers a memory primitive that the developer wires to a store. The StoreMemory action in AgentFlow lets developers explicitly save previous user queries, model outputs, intermediate decisions or system states for recall in later interactions, retrieved into subsequent prompts. A SessionId passed by the consumer app writes conversation state to an ODC entity in the customer's own database, and a shared memory pattern coordinates multiple agents. OutSystems does not describe a retention period, purge path or review and deletion surface for stored memory.

Sourcesuccess.outsystems.com agentic apps in ODC pageread 2026-09-29

Human approval set on a tool makes a visual agent pause before calling it and ask the user to confirm, optionally letting the person reviewing edit the tool inputs (not inside For Each, Parallel or Reflection sub sequences), and guardrails cover PII, prompt injection, toxicity, topic boundaries and bias.

Sourcedoc.dataiku.com/dss/latest/agents/tools/human-approval.htmlread 2026-09-27

Agents can be held for human oversight on high stakes or low confidence decisions. A Human activity element placed in the workflow pauses execution until an assigned user or role acts, the assigned person sees the agent's output on a task screen and clicks Approve or Reject, and that action raises an event closing the activity so the workflow resumes on the human's decision.

In one example, a vendor approval workflow pauses after the agent generates a risk report. A validation branch can route a rejected result back to a prior step to rerun the agent with feedback. Separately, Agent Guardrails are enforced in the runtime and cannot be bypassed once enabled, covering prompt attack blocking, PII masking and harmful content filtering.

Sourcesuccess.outsystems.com building agentic workflows and agentic patterns documentationread 2026-08-31

Dataiku offers SSO through SAML, OIDC, LDAP and Kerberos, MFA, user profiles, per project group and user permissions, connection security, user secrets and an audit trail. Its trust page lists ISO 27001:2022, ISO 27701, ISO 42001 and ISO 9001 certification, SOC 1 and SOC 2 Type II for Dataiku Cloud, and a HIPAA report.

Sourcedataiku.com/legal/trustread 2026-09-27

OutSystems runs a public Trust Center at security.outsystems.com with certification records and security advisories. Attestations include SOC 2 Type II, ISO 27001 and ISO 22301 certified by BSI, ISO 27017 and ISO 27018 for cloud security, and CSA STAR self assessment, with OutSystems Developer Cloud separately holding SOC 2 attestation evaluated by Baker Tilly and SOC 1 Type 2 attestation.

An identity service provides role based access to screens, data and logic flows with administrator controlled permissions, single sign on and external authentication, and a formal logical access procedure. Agent Guardrails are enforced in the runtime rather than at the application layer, detecting and blocking prompt attacks, masking or blocking PII before data reaches a model provider, and filtering harmful content.

Sourceoutsystems.com evaluation guide security and certifications pages, security.outsystems.com Trust Center, ODC attestation announcementsread 2026-08-31

Every LLM Mesh call returns a nested trace of the agent's LLM calls, guardrail steps, inputs, outputs, tokens and cost, viewed in the Trace Explorer (tree, timeline and explorer views), and the evaluation trajectory explorer replays tool calls in order; an audit trail records user actions.

Sourcedoc.dataiku.com/dss/latest/agents/tracing.htmlread 2026-09-27

Distributed traces treat an agent as a filterable asset in its own right, filterable by stage, asset, element type, trace status, total duration, date and time, and user, retained up to four weeks and queryable in two week intervals, and spans are aggregated so an agent's execution can be read both as its own trace and as a span inside a calling app's trace.

Logs carry Error, Warning and Information severities, are generated automatically or through LogMessage actions developers place, and link to related logs and activities for root cause work. Agent Evaluations return quality scores with full execution traces.

Sourcesuccess.outsystems.com monitoring and troubleshooting apps documentationread 2026-08-31

Dataiku runs as fully managed Dataiku Cloud or is installed on the customer's own infrastructure (Linux, macOS, experimental Windows) with hybrid options through sales, with installation, cloud and container deployment guides; locally running Hugging Face models keep inference in house.

Sourcedataiku.com/product/get-startedread 2026-09-27

With ODC self hosted deployment, customers run their apps, data and AI agents in their own private cloud, public cloud, or on premises Kubernetes infrastructure while OutSystems continues to manage platform services, with single tenant runtime provisioning from the ODC Portal and a secure pull mechanism for updates.

OutSystems aims this at customers who need absolute control over hosting environments for data residency, security or regulatory compliance mandates. A secure private gateway connects ODC to on premises databases, APIs and enterprise systems without moving data.

A self hosted provisioning repository supporting Kubernetes and OpenShift is published on OutSystems' GitHub, and the OutSystems 11 platform runs on self managed infrastructure across multiple cloud providers and on premises.

Sourceoutsystems.com ODC self-hosted product update, developer cloud product page and deployment evaluation guide, github.com/OutSystems/outsystems-hybrid-provisionread 2026-08-31

Dataiku Solutions include named GenAI and agent solutions for separate jobs, Dynamic Selling Assistant, Supplier Management Assistant, Clinical Trial Intelligence Assistant, Medical Entity Extraction Assistant and Agentic Insights; Agent Hub distributes an organization's own governed agents and prompt library.

Sourceknowledge.dataiku.com/latest/solutions/index.htmlread 2026-09-27

Agent Workbench shipped an agent marketplace at general availability alongside MCP support. Forge, a public browsable catalog of numbered, individually addressable and documented components, installs components into an app and exposes their server actions to agents, OmniDoc2MD among them.

Agent creation is template-driven, starting from defined components for grounding data, prompt, model and tools, and a Mentor capability in beta generates a full agent configuration including instructions, tools and model settings from a plain-language description with conversational refinement. OutSystems launched its first end to end Agentic Industry Solution, a Banking Solution for Loan Origination, at ONE, as the first of a family.

Sourceoutsystems.com Agent Workbench GA announcement, ONE 2026 announcement and Q2 2026 product update, outsystems.com Forgeread 2026-08-31

Administrators define LLM Mesh connections to OpenAI, Azure OpenAI, Anthropic, AWS Bedrock and SageMaker, Microsoft Foundry, Google Vertex, Mistral, Cohere, NVIDIA NIM, Snowflake Cortex, Databricks foundation model APIs and locally running Hugging Face models, and agents pick from them, with cost control and rate limiting documented.

Sourcedoc.dataiku.com/dss/latest/generative-ai/llm-connections.htmlread 2026-09-27

In Agent Workbench a model is connected once and reused across agents, with support at general availability for AWS Bedrock, Azure OpenAI and Anthropic, plus Gemini, Cohere, Mistral, Databricks, AI2 and IBM watsonx, custom built models on Vertex AI and open source models on Hugging Face. An expanded Amazon Bedrock partnership gives teams access to more than one hundred large language models through a single set of credentials, and models can be swapped as pricing or performance shifts without rewriting underlying logic. OutSystems pitches the platform as insulating customers from LLM lock in.

Sourceoutsystems.com Agent Workbench GA announcement, Q2 2026 product update and ONE 2026 platform announcementread 2026-08-31

Agents are exposed as A2A servers at a per agent public API endpoint with well known URI discovery and bearer API key auth, in JSON-RPC and SSE modes, beside a documented MCP server, the public REST API, the Python API and LangChain integration on developer.dataiku.com, and a plugin system.

Sourcedoc.dataiku.com/dss/latest/agents/a2a.htmlread 2026-09-27

ODC publishes REST APIs for the platform itself, authenticated through API clients using OpenID Connect over OAuth 2.0 client credentials, with twelve named references (User and Access Management, Portfolio v2, Build Operations, Native Mobile Build, Deployments, Dependency Management, Asset Repository, Asset Configurations, Environment Configurations, Code Quality, External Library Generation Service and Subscription). Agents are also exposed through REST Expose as inbound endpoints, and OutSystems has announced Agent Experience as a layer of A2A and MCP services, without publishing addresses for those services.

Sourcesuccess.outsystems.com ODC REST APIs pageread 2026-09-29

The Evaluate Agent recipe scores answers and the path taken, with ordered and unordered tool call checks, RAGAS based LLM judges, BERTScore and custom metrics, stores results for side by side version comparison, and Agent Review lets builders and SMEs run test cases against an agent version with human verdicts overriding the judge.

Sourcedoc.dataiku.com/dss/latest/agents/evaluation.htmlread 2026-09-27

Agent Evaluations validate agent logic against a Golden Dataset of inputs, expected outputs and expected tool calls, returning quality scores and full execution traces, with an automatic judge scoring results rather than requiring manual inspection.

OutSystems pitches Evaluations as replacing manual one off testing with a repeatable validation process, and advanced agent evaluations, launched at ONE, as validating the outcomes agents are responsible for the way unit tests validate code.

ODC also offers debugging with breakpoints, step by step execution and runtime value inspection, and automated testing strategies for application delivery.

Sourceoutsystems.com ODC Agent Evaluations product update and ONE 2026 announcement, success.outsystems.com agentic development documentationread 2026-08-31

The agent tool list covers APIs, MCP, knowledge search, code and an LLM side web search tool, with no browser or computer use tool documented.

Sourcedoc.dataiku.com/dss/latest/agents/index.htmlread 2026-09-27

Agents act through server actions selected by the model, through prebuilt connectors to ERP, CRM, databases and legacy systems, and through Model Context Protocol, all of which are programmatic interfaces. OutSystems does not describe browser control, navigation, form filling, screen interaction, visual grounding or desktop automation, and makes no claim in that territory. The platform's legacy modernization path analyzes COBOL and Lotus Notes systems and rebuilds their functionality as new applications through code transformation.

Sourcesuccess.outsystems.com agentic apps and AI agent actions documentation, outsystems.com ONE 2026 announcementread 2026-08-31