Dataiku vs OutSystems (2026)

Stands at F, re-based onto the AI agent actions documentation, which describes the tool-calling half far more precisely than the July basis did. TOOL CALLING IS DOCUMENTED AS GENUINE AUTONOMOUS SELECTION, not a scripted call: the model DECIDES WHICH ACTION OR SEQUENCE OF ACTIONS IS APPROPRIATE by analysing the prompts and the stated purpose of each exposed action, and may call one action, a sequence, or the same action repeatedly. Repeat invocation is the detail that indicates a real agent loop rather than a single-shot function call. The integration shape is the distinctive part and worth carrying to comparison pages. Rather than a fixed connector catalogue, ANY SERVER ACTION BECOMES A TOOL. Twenty-five years of enterprise integration built for the low-code platform, ERP, CRM, databases, legacy systems, becomes agent-callable without a separate agent-connector programme. That is a different trade from the hundred-connector platforms in this lane: coverage is bounded by what the customer has already integrated rather than by what the vendor has built. Forge extends it further, and the documentation's own worked example is instructive: add the OmniDoc2MD component and expose its conversion action so the agent can ingest Word, PDF, spreadsheet, presentation and HTML content. MCP adds the inbound direction for enterprise systems and tools; the outbound exposure of OutSystems agents to external assistants is credited on Ext and not counted twice here. One documented constraint recorded rather than penalised: action calling and structured output cannot be used on the same agent call.

Stands at F on documentation rather than the product-page claim the July basis used. The substantive finding is that MULTI-AGENT COORDINATION IS DOCUMENTED AS NAMED ARCHITECTURES, not asserted as a feature. A shared-memory pattern where agents read and write common state and communicate IMPLICITLY BY MODIFYING IT rather than by messaging each other, and a network pattern of autonomous peers interacting in a graph-like structure WITHOUT A CENTRAL CONTROLLER. Documenting decentralised coordination is a stronger claim than the orchestrator-and-workers shape most vendors in this lane describe, because it admits emergent behaviour and then supplies patterns for governing it. THE ARCHITECTURE WORTH CARRYING is the split between agents and workflows, which the documentation states plainly: agents provide reasoning and autonomy, workflows provide STRUCTURE, STATE MANAGEMENT AND HUMAN OVERSIGHT. Automatic activities call agents, human activities gate them, validation branches loop them back with feedback. Deterministic control flow surrounds non-deterministic reasoning rather than being replaced by it, which is the right shape for processes where a wrong step has consequences and is the same split maisa and sema4-ai make. Unlike several Full cells in this lane, the control-flow vocabulary is fully present: branching, looping back, sequencing, hierarchy and parallel side-by-side execution are all documented. Agentic Enterprise Orchestration, the current generation launched 3 June 2026, adds semantic search and the A2A connector to this layer.

P>F. The July basis said a broad real-time or multi-channel trigger framework was not the emphasis, which was a failed pass rather than a finding: the documentation names the trigger classes in a single sentence. The decisive line is in the monitoring documentation, and it is about agents specifically rather than apps generally: filtering traces by an agent returns flows TRIGGERED BY AN EVENT HANDLER, A TIMER, OR A REST EXPOSE. That is three classes stated as first-party fact, and it is a stronger form of evidence than a marketing page listing triggers, because it is the platform's own observability surface describing what it actually observes. Each class is separately documented. Timers are background scheduled jobs with cron-like schedule syntax, created in the Events tab against a server action, and forceable on demand through the Wake action or the ODC Portal. Events run through the platform's event-driven architecture. REST Expose makes the agent an inbound HTTP endpoint, which is the webhook class. The channel side is covered separately and does not carry the grade: agents publish as service actions that any app in the tenant consumes, so they embed into web and mobile interfaces, and the Agent Experience layer exposes MCP and A2A services so an assistant the customer already uses can reach them. Schedule, event, inbound API and embedded application is the full spread this axis asks for.

Stands at F and sits on the correct side of the Knowledge line, which asks whether a maintained retrieval structure exists rather than per-request assembly. THE ENTERPRISE CONTEXT GRAPH IS THE MAINTAINED STRUCTURE, and it is an unusual one for this lane. Most vendors' knowledge layer is a document index over content the customer uploads. This is a graph over the customer's own APPLICATION ESTATE, entities, relationships, roles and business logic, derived from twenty-five years of building software around business intent rather than code alone. It stays current because it is generated from the running estate rather than synced from a corpus, which is the same structural argument that put sema4-ai's zero-copy access at Full: there is no stale index because there is no index to go stale. The vendor's own framing is that the graph grounds ANY CODING TOOL OR LLM in a shared operational knowledge layer, which is a claim about it being a substrate rather than a feature, and the Agent Experience layer exposing it over MCP is consistent with that. Grounding data is the per-agent half: agent configuration begins by connecting internal databases or external APIs so responses rest on current factual information. Semantic search shipped 3 June 2026 with Agentic Enterprise Orchestration, after the record was researched. Document ingestion is documented through Forge conversion components handling Word, PDF, spreadsheets, presentations, HTML and plain text with structure preserved and images extracted. One limit: no chunking, embedding or retrieval configuration was reached on a page this pass.

P>F. The July basis said a distinct persistent agent memory was not the emphasis, which the documentation contradicts directly: STOREMEMORY is a named agent action, not an inferred capability. Three documented structures carry it. StoreMemory saves the conversation or generated information FOR FUTURE REFERENCE, explicitly to enable multi-turn or personalized interactions. SessionId is generated and managed by the consumer app and passed to the agent to maintain conversational context ACROSS INTERACTIONS, so session identity is a first-class parameter rather than an implementation detail. And the agentic patterns documentation describes a shared memory pattern where agents read from and write to shared state, communicating implicitly by modifying it, which is memory operating as the coordination substrate for a multi-agent system. The documentation also points to a dedicated state persistence section, which is the vendor treating this as a platform concern rather than a per-app problem. THE HONEST LIMIT, and the reason confidence is medium rather than high: conversation state is written to an ODC entity the customer owns and queries, and no automatic extraction, summarisation or learning-from-correction layer is documented. The customer wires the store using supplied primitives. I record that this sits adjacent to the sema4-ai reasoning, where accumulation was described as the condition that WOULD take a cell to Full. I read that as sufficient rather than necessary: the axis measures memory and state across sessions, and a named memory action plus session identity plus documented state persistence is a shipped memory surface, not an assembled one. Flagged in the report for consistency rather than decided quietly. I could not open the state persistence page itself; success.outsystems.com blocks automated retrieval, so the grade rests on the agent creation and patterns pages.

Stands at F, and this is now graded on a shipped mechanism rather than the governance prose the July basis rested on. The old basis cited built-in oversight and strict access controls, which is exactly the essay language section 7 says not to grade from; the documentation supplies the actual engine. THE MECHANISM IS SPECIFIC AND IT IS THE VENDOR'S OWN. A Human activity element placed in the workflow PAUSES EXECUTION until an assigned user or role acts. The reviewer sees the agent's output on a task screen and clicks Approve or Reject. That action raises an event that closes the activity and resumes the workflow on the human's decision. Assignment to a role rather than an individual is what makes it operable at enterprise scale. THE 30 AUGUST RULING IS CLEARED PROPERLY, and it is worth being explicit given how many near-misses it has caught this lane. This is not a gate the customer already owns: the task screen, the pause, the assignment and the resume event are all OutSystems' own workflow surface, documented specifically for agent oversight. It is the opposite of the draft-pull-request case. The rejection path is stronger than a simple stop: a false branch routes back to a step before the agent call, rerunning it with feedback or updated context, so the human's judgment re-enters the agent's input rather than merely halting it. Runtime guardrails sit underneath as a second, non-optional layer, and the fact that agents cannot bypass them once enabled is the part that distinguishes them from prompt-level instructions.

Stands at F, and confidence rises from medium to high because the July basis asserted enterprise-grade governance without naming a single certification, control or page. Both halves of the conjunction are now documented and independently verifiable. THE ATTESTATION HALF IS UNUSUALLY DEEP. SOC 2 Type II, ISO 27001 and ISO 22301 from BSI, ISO 27017 and ISO 27018 for cloud-specific controls, CSA STAR self-assessment, and separate SOC 2 and SOC 1 Type 2 attestations for ODC specifically, the SOC 2 evaluated by Baker Tilly. Named audit firms and a live Trust Center at security.outsystems.com put this well past the hedge ladder in section 7, where the top rung is Type I with Type II in progress. THE CONTROL HALF IS EQUALLY REAL. An identity service with role-based access to screens, data and logic flows, administrator-controlled permissions, SSO and external authentication. WORTH CARRYING FOR THE AGENT LANE: Agent Guardrails are built into the RUNTIME rather than bolted on, so agents cannot bypass them once enabled, and PII masking happens BEFORE DATA REACHES A HYPERSCALER OR FOUNDATION MODEL PROVIDER. Most vendors in this lane place safety controls in the application layer where a determined agent path can route around them, and the egress-boundary placement of the masking is the specific detail a regulated buyer would ask about. Per the 30 August ruling, sovereign and hybrid deployment properties are graded on Dep and are deliberately not counted again here.

Stands at F, re-based from an unnamed docs citation onto the monitoring documentation, and it clears the reporting-versus-auditing line properly rather than by assertion. THE DECISIVE PROPERTY IS THAT AN AGENT IS A FIRST-CLASS FILTERABLE ASSET IN THE TRACE SYSTEM, not an opaque step inside an application trace. Traces filter by stage, asset, element type, trace status, total duration, date and time, and user. That means a specific agent's behaviour can be isolated across an estate, which is the question an operator actually asks when something went wrong. The span model is the detail worth carrying. The documentation distinguishes traces where the agent is the ORIGIN from traces where the agent RUNS AS A SPAN INSIDE ANOTHER APP'S TRACE. Both views exist, which means a multi-agent chain can be reconstructed end to end rather than as disconnected fragments, and this is the why rather than only the what. Retention is stated rather than vague: four weeks back, filterable in two-week intervals. Naming a retention window is uncommon in this lane and is the thing a compliance buyer needs. Logs sit alongside with three severities, automatic generation plus developer-placed LogMessage, and a related-logs-and-activities panel for root-cause work. Evaluations contribute full execution traces on top. One limit: export to a customer's own SIEM or observability stack was not documented on any page reached this pass.

F RESTORED ON A PAGE I HAD NOT REACHED, after the 31 August ruling correctly found my earlier basis did not clear the bar. Recording both halves, because the correction matters more than the outcome. WHAT WAS WRONG WITH THE EARLIER BASIS. It rested on three facts and none of them carried. A defined organization region with a warning about processing leaving it is a fixed region, disclosed but not selected, and disclosure is not residency. PII masking before data reaches a hyperscaler is a security control already credited on Sec, so counting it here was one fact working two axes. The AWS EU sovereign expansion is explicitly still to come, which section 7 excludes as roadmap. That left self-managed deployment as the only load-bearing fact, and I had not verified it. All three are removed from the basis. WHAT ACTUALLY CARRIES IT, published 31 March 2026 and missed by both the July build and my first pass: ODC SELF-HOSTED DEPLOYMENT. Customers run apps, data AND AI AGENTS in their own private cloud, public cloud, or on-premises Kubernetes infrastructure while OutSystems continues to manage the platform services. The vendor states the driver plainly, customers needing ABSOLUTE CONTROL OVER THEIR HOSTING ENVIRONMENTS for data residency, security or regulatory mandates. THE DETAIL THAT MAKES IT REAL RATHER THAN A DEDICATED-TENANT REBRAND: single-tenant runtime provisioning directly from the ODC Portal, and a SECURE PULL MECHANISM for updates, meaning the customer's environment reaches out rather than accepting inbound connections. That is the architecture an air-gapped or tightly firewalled buyer requires, and it is the difference between hosting location and hosting control. The product page states the split cleanly: platform services managed by OutSystems, runtime applications running in customer infrastructure. The agent runtime being named in scope is what makes this an agent-platform deployment fact rather than an inherited application one. One conflict recorded and resolved: a third-party pricing analysis from May 2026 states ODC runs exclusively in OutSystems-managed cloud. It is an aggregator, it is not evidence under section 7, and it is contradicted by the vendor's own product update two months earlier.

Stands at F under the 31 August bar, which asks whether the customer receives packaged assets ready to adopt by any route and treats a browsable catalogue as evidence rather than as the definition. FOUR ROUTES QUALIFY, WHICH IS MORE THAN THE BAR REQUIRES. An agent marketplace shipped with Agent Workbench general availability. Forge is a public catalogue of numbered, individually addressable, documented components that install into an app, and the documentation's own example of extending an agent is adding one. Agent creation is template-driven from defined components. And a Mentor beta generates a complete agent configuration, instructions, tools and model settings, from a plain-language description. THE INDUSTRY SOLUTIONS LINE IS THE STRONGEST FORM AND THE NEWEST. The Banking Solution for Loan Origination launched at ONE on 3 June 2026 is an end-to-end solution for a named business process, described as the first of a family of Agentic Industry Solutions. That is a finished agentic system for a vertical, not a starting template, and it is the closest thing in this lane to what a buyer means by a pack. On the 32 percent finding in the 30 August brief, that only a third of Full cells index-wide name a browsable catalogue: this record names one that predates the agent era and is genuinely browsable, so it sits in the defensible third rather than the majority that assert templates without a surface. For the record, the Mentor generator is beta and is corroborating rather than load-bearing.

Stands at F and is close to the strongest form this axis takes: the customer chooses the model, the choice is cheap to revise, and the vendor makes non-lock-in an explicit product position rather than a side effect. The provider list is genuinely broad rather than a hyperscaler plus a token second option: Bedrock, Azure OpenAI, Anthropic, Gemini, Cohere, Mistral, Databricks, AI2 and IBM watsonx, plus CUSTOM-BUILT MODELS ON VERTEX AI and OPEN-SOURCE MODELS ON HUGGING FACE. The last two matter most, because a customer's own fine-tuned model and an open-weights model are the two cases that separate real flexibility from a curated menu. THE ARCHITECTURAL DETAIL THAT EARNS FULL RATHER THAN A LONG LIST: a model is connected once and reused across agents, and models swap as pricing or performance shifts WITHOUT REWRITING UNDERLYING LOGIC. Model choice is a configuration concern separated from agent logic, so the choice stays revisable after deployment instead of being frozen at build time. The Q2 2026 Bedrock expansion adds more than one hundred models through a single set of credentials, which collapses the procurement friction that normally makes multi-provider support theoretical. Per section 7 this is model flexibility and not an Ext credit; the MCP and A2A surfaces run the other direction and are graded there.

Stands at F and is now one of the better-evidenced Ext cells in the lane, against a July basis that listed capabilities generically without naming MCP, A2A or a single endpoint type. Mike's 30 August bar asks whether the platform is callable from outside through a documented API or SDK. It is, three ways: REST Expose makes an agent an inbound HTTP endpoint, published agents become service actions consumable by any app in the tenant, and the Agent Experience layer exposes MCP and A2A services. THE AGENT EXPERIENCE LAYER IS THE MATERIAL CHANGE SINCE THE RECORD WAS BUILT, launched at ONE on 3 June 2026 and missed by a build dated 7 July. It is a platform layer exposing a suite of A2A and MCP tools and services, with the first services live for agentic coding, publishing and platform extensibility. The direction is the Ext direction, not the Model one: the customer's own assistant reaches into the OutSystems estate. MCP support itself is older than the record realised, shipping with Agent Workbench general availability in September 2025. The A2A connector is the piece worth carrying to comparison pages, because it runs both ways: external agents integrate into an OutSystems agentic system, and OutSystems agents participate in someone else's. Few vendors in this lane document both directions. Native integration with Claude Code, Codex, Cursor and Kiro is corroborating rather than load-bearing, but it is unusual for an enterprise platform to treat outside coding agents as first-class clients.

Stands at F and clears the 31 August Eval bar decisively, which asks for a readable comparable result rather than a quality gate with no readable output. Evaluations return QUALITY SCORES plus full execution traces against a Golden Dataset specifying inputs, expected outputs AND EXPECTED TOOL CALLS. Grading the tool calls as well as the text is the part that matters: it catches an agent that reached a plausible answer by the wrong route, which output-only scoring cannot see. The same distinction earned sema4-ai its Full earlier in this batch. THE AUTOMATIC JUDGE IS WHAT MAKES IT REPEATABLE RATHER THAN CEREMONIAL. A harness a human has to read is run once before launch; a scored one is run on every change, and the vendor's own framing is that Evaluations REPLACE MANUAL, ONE-OFF TESTING WITH A REPEATABLE VALIDATION PROCESS. The vendor's unit-test analogy is apt rather than marketing: a golden dataset with expected outputs is a regression suite, and the comparison it supports is the before-and-after of a prompt, tool or model change, which is the question this axis exists to ask. Debugging is documented separately with breakpoints, step-by-step execution and runtime value inspection, so the loop covers both measuring that something broke and finding where. Advanced evaluations shipped at ONE on 3 June 2026, after the record was researched but before it was built.

Stands at N, and confidence rises to high because the architecture settles it rather than leaving it unretrieved. The July basis was correct and its wording, that the platform does not drive a browser as a first-class capability, was one of the few honest cells on this record. Everything documented runs through programmatic paths: server actions selected by the model, prebuilt connectors to ERP, CRM, databases and legacy systems, and MCP. There is no browser, no screen and no interface being operated, and the vendor makes no claim in that territory anywhere. ONE NEAR-MISS WORTH NAMING AND REFUSING, because it is the kind that has been credited elsewhere in the index. Legacy Modernization Services analyse COBOL and Lotus Notes systems and rebuild their functionality as governed applications. Reaching into systems that predate APIs sounds like the Comp case, but it is the opposite: the platform reads and transforms source and structure, then replaces the system. Comp is non-zero only when an agent OPERATES software because no programmatic interface exists. Modernization removes the need to operate the old interface rather than operating it. The absence is structural and correct for what this is. A platform whose knowledge layer is a graph over the customer's own application estate has programmatic access to those systems by construction, so driving a screen would be a strictly worse route to data it can query. With orby-ai watchlisted and sema4-ai at Full on Robocorp desktop libraries, this record is a clean negative reference point for the lane.