OutSystems Agent Workbench

Spectrum 0.2 beta · Public tier · read 2026-11-10 · not reviewed by its vendor yet

Builds, runs and governs AI agents (agentic apps) on OutSystems Developer Cloud (ODC): agents call server actions, MCP tools and external A2A agents, with guardrails, evaluations, traces and workflow human-in-the-loop steps.

What each level rests on

This card rates a product as its vendor documents it, not anyone's deployment of it. A feature the operator has to turn on counts, and is marked as available. Levels are computed from the evidence below by published rules. A dimension with no evidence reads "not yet evidenced".

Autonomy: level 4, Orchestrates

What does the agent do on its own? Runs multi-step workflows across systems or coordinates other agents; a person sets goals and reviews outcomes. This level holds with: Hitl workflow; Agent workflow; A2a connection. As shipped, with nothing turned on, it is not yet evidenced.

• I10 It works across several domains or systems. The product FAQ says agents connect to APIs, databases, ERP, CRM and legacy systems and orchestrate workflows across platforms. "Agents orchestrate workflows across platforms while maintaining security and governance." Source (read, documented, captured 2026-11-10T15:46Z, sha256 b8d83b34ccd1be37) · π Moderate · S1·R5·A1 ×3
• I12 Several agents coordinate inside the organisation. Workflows chain several agents, each with its own lifecycle, passing one agent's output to the next. "With this pattern, you can chain multiple agents together sequentially." Source (read, documented, available, captured 2026-11-10T15:46Z, sha256 409bddbad33512f9) · π Moderate · S1·R5·A1 ×3
• I13 Agents operate across organisational boundaries or disparate stacks. Agentic apps can delegate tasks to external agents, outside ODC, over the A2A protocol. "allows your ODC agentic apps to interact and collaborate with external agents" Source (read, documented, available, captured 2026-11-10T15:46Z, sha256 7bc6c8084a21d0cf) · π Moderate · S1·R5·A1 ×3
• I18 The agent works independently; the person approves critical decisions only. In a workflow, agents run as automatic activities completed without human intervention, and a Human activity assigned to a user or role pauses the workflow until that person approves or rejects the agent's output. "allowing the workflow to resume based on the human's decision" Source (read, documented, available, captured 2026-11-10T15:46Z, sha256 409bddbad33512f9) · π Moderate · S1·R5·A1 ×3

This evidence touches 14 provisions: Salesforce, Microsoft, OWASP, AAGMM, EAAMM.

Reach: level 3, Open protocol

How connected is it? A published agent interface with documented authorisation. This level holds with: MCP connection; A2a connection. As shipped, with nothing turned on, it is level 2, Integrated.

• I14 There is a published protocol or communication layer between agents. The A2A connector implements version 0.3.0 of the published Agent2Agent protocol, reading the remote agent card. "ODC A2A connector supports only A2A 0.3.0 protocol version" Source (read, documented, available, captured 2026-11-10T15:46Z, sha256 305b5bf8b523b55b) · π Moderate · S1·R5·A1 ×3
• I25 It asks for narrower access than the whole account. The builder grants each agent a set of server actions it may call; with none granted, the agent relies only on the AI model's knowledge. "You provide the agent with a set of available actions." Source (read, documented, captured 2026-11-10T15:46Z, sha256 b9307656fbaee2ea) · π Moderate · S1·R5·A1 ×3

This evidence touches 8 provisions: Salesforce, AAGMM, MCP authorization.

Next unlock: level 4, Ecosystem. Public evidence that others build on it: third parties publish integrations, skills or agents for it.

Control: level 2, Configurable

Can people stay in command? Approval modes, permission scopes and roles are documented settings. This level holds with: Hitl workflow. As shipped, with nothing turned on, it is level 1, Stated.

• I38 High-impact actions need a person's approval. The human-in-the-loop pattern stops at a high-stakes trigger until a person verifies or decides. "alerts a human user to provide verification or a decision before proceeding" Source (read, documented, available, captured 2026-11-10T15:46Z, sha256 ad4921b706fb6cc0) · π Moderate · S1·R5·A1 ×3
• I25 It asks for narrower access than the whole account. The builder grants each agent a set of server actions it may call; with none granted, the agent relies only on the AI model's knowledge. "You provide the agent with a set of available actions." Source (read, documented, captured 2026-11-10T15:46Z, sha256 b9307656fbaee2ea) · π Moderate · S1·R5·A1 ×3

This evidence touches 11 provisions: OWASP, EU AI Act, AAGMM, NIST AI RMF, Salesforce, Microsoft, MCP authorization.

Next unlock: level 3, Observable. Public evidence that its actions are logged; and that it is monitored in operation.

Openness: level 2, Documented

Does it show its work? Public documents on how it works, its data, its models and its limits.

• I22 Its limits and known failure cases are documented for those who use it. The guardrails page documents where guardrails do not apply: files, some languages in basic-coverage regions, and two unsupported regions. "Prompts or responses in other languages can bypass safety filters." Source (read, documented, captured 2026-11-10T15:46Z, sha256 adaf537c0b182a67) · π Moderate · S1·R5·A1 ×3
• I58 The data it uses is described. The docs describe agent grounding data: internal aggregates, search services and external tools. "Sources can include internal aggregates, search services (RAG), or external tools." Source (read, documented, captured 2026-11-10T15:46Z, sha256 4bebd8d78dce8555) · π Moderate · S1·R5·A1 ×3
• I61 The models and third-party components it uses are listed. The docs list the supported model providers (OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, Gemini and others) and the trial models. "ODC supports connections to AI models from multiple providers" Source (read, documented, captured 2026-11-10T15:46Z, sha256 0727bf95007ca798) · π Moderate · S1·R5·A1 ×3

This evidence touches 9 provisions: NIST AI RMF, EU AI Act, AAGMM, EAAMM, Salesforce, OWASP.

Next unlock: level 3, Measured. Public evidence that test or evaluation results are published.

How far to trust this card

Status: Draft. Draft means the facts were found in public pages and reviewed by agents. The company has not confirmed them and no customer has gone on record. FACT-Ray says how far each fact can be trusted: who says it (Source), how firmly it is recorded (Record) and who stands behind it (Accountability). It never changes a level.

• The facts: 22 facts: 22 π Moderate. Each is written beside its evidence above.
• The chain: found by an agent, captured and hashed, checked by 2 other agents (same model family, shared repository); no human has signed it.
• Who: produced by gamma.studio@web-summit.dinis; checked by gamma.review@web-summit.dinis and gamma.approve@web-summit.dinis. The actors.
• What would raise it: the company's own file at /.well-known/fact-ray.json (Source 2), a customer's statement of use (Source 3), or a named person signing the card (how).

The standards constellation

One piece of public evidence can speak to several frameworks at once. "Touched" never means "complies". The share is of the provisions that can be read from outside.

The 61 provisions lit

Each line item, with the facts that reach it. "Evidenced" is a link of the kind same: the fact is what the provision asks. "Partly" is a link of the kind part: the fact is one piece of it. "Its absence" points the other way. Each provision has its own page with its text and its place in the source.

Upper layers

The public card is the front door. The same evidence, read with the governance rules, is this agent's GAMMA-Ray rating: freedom, governance, the level held, the grant and a work order. That layer is written for an operator.

The vault

Every source behind this card was captured and hashed into a public sgit vault (ivinvfhj): captures.json lists each URL with the SHA-256 of what was fetched and when. Read key:

sgit_public_read_a787d5ebd3b7fca03039c35a5d95311a66cc5581cbc41eec0ad3ab0c46272feb:ivinvfhj

With sgit: sgit clone sgit_public_read_a787d5ebd3b7fca03039c35a5d95311a66cc5581cbc41eec0ad3ab0c46272feb:ivinvfhj.