DataFab //  artificial proprietary intelligence T+00:00:00 ISO/IEC 27001:2022 · SOC 2 Type II — certified Raw egress  0 B 

DataFab  /  Platform  /  Knowledge & Agentic Studio

Layer 2 · the builder

Inside the Studio, words to a governed agency.

The build surface of DataFab — where the enterprise composes, tests and publishes its own governed agencies, from its own documents, in its own language, and on its own schemas. Not a bench of specialists, and not a vendor engagement for every new workflow.

You build themIn words, not codeGoverned by construction

01 · the builder

Everything needed to compose governed execution.

Build, test and publish data-driven agents, design workflows, define widgets and orchestrate multi-agent solutions. You build the agencies — on your schemas, in your language. The Studio does the wiring, the testing and the governance.

Domain discovery

Extract schemas directly from the documents you already have.

Data-driven agents

Schema-bound processing units — the unit of execution.

Visual pipeline builder

Drag-and-drop composition with live validation, no code required.

Planning from language

Build an agent from a plain-language description of the outcome.

Chain of agents

Orchestrate sequential work with human gates wherever judgement belongs.

Graph of agents

Non-linear orchestration — branching, fan-out and fan-in. Planned.

Widgets

Visual interface components the analyst can drive directly.

Datasets

Structured, schema-bound data available to every agent.

Utilities

Reusable API and agent components, published for others to compose.

MCP integrations

Connect — and author — your own connectors.

Unified tool catalogue

Every built-in node, fabric service and connector in one searchable place.

Template library

Pre-built workflows to start from rather than a blank canvas.

The approach

From words to a governed agency.

Describe the outcome. The Studio derives the schema, plans the pipeline, and hands you something to test — built from what you already have. The wiring, the schema-mapping and the guardrails are built for you.

01

Describe

Say what the agency should do, in plain language. No query language, no orchestration syntax.

02

Derive

Schemas are extracted from your own documents — specifications, policies, contracts, forms.

03

Plan

A query plan and pipeline are generated for you, and imported straight into the visual builder if you want to shape it.

04

Test

Run it in an isolated sandbox against sample files, and validate output against the schema before anything is published.

05

Publish

Set it live, scoped and governed, with runtime placeholders mapped and access permissions applied — reusable by anyone entitled to it.

Why compose rather than buy

Vanilla stands beside the estate. An agency stands on it.

A vendor product reaches one system at a time, and the next deployment starts again. An agency composed here is built on the resolved estate, so it applies everywhere the moment it is published.

Fig. 02b  //  beside the estate, or on it◆ one at a time, or all at once
THE SAAS WAYbeside the estateONE AT A TIMECRMDMSERPDWHSTREAMREGCASELEGACYVendor Aintegration 1Vendor Bintegration 2Vendor Cintegration 3EACH ONE REACHES A SINGLE SYSTEM · THE NEXT STARTS AGAINTHE STUDIO WAYon the resolved estateALL AT ONCETHE RESOLVED ESTATE · ONE GOVERNED GRAPHCRMDMSERPDWHSTREAMREGCASELEGACYYOUR AGENCYcomposed in the StudioALREADY RESOLVED BENEATH IT · IT APPLIES THE MOMENT YOU PUBLISHTHE DELTA — YOU STOP RENTING PROCESS AND START ACCUMULATING AN ASSETWHAT IT KEEPS COSTINGWHAT YOU KEEP INSTEADA licence per workflow, renewed foreverAn integration project per system, per vendorA vendor engagement for every new use caseThe same capability bought again for the next entityThe resolved estate — built once, reused by everything afterThe encoded workflow as an owned asset, not a renewed engagementEvery next agency cheaper and faster than the lastSchema, graph and agencies exportable — you can leave with themNOT ONE INTEGRATION AT A TIME · ALL OF IT, AT ONCECOMPOSED IN YOUR WORDS · ON YOUR SCHEMAS · OWNED BY YOU

You stop renting process and start accumulating an asset. The full argument, including what each route keeps costing and what it leaves you owning: Don’t buy vanilla →

Two hands on the same tool

A business user and a platform engineer, on one artefact.

The Studio meets each where they are strongest — and both routes converge on the same governed artefact: schema-bound, sandboxed, human-gated and audited.

Business user · tech-novice

Build it in language.

  • Describe it in words — planning produces a draft agent
  • Upload your documents — domain discovery produces the schemas
  • A sentence becomes a pipeline you can see and edit
  • Start from a template of pre-built workflows

Expert · full control

Compose it with every node yours to shape.

  • Compose in the visual builder with live validation
  • Write custom logic and scripts inside the query plan
  • Author custom connectors in MCP Studio
  • Define schemas and query plans directly in JSON
  • Debug step by step with breakpoints and traces

02 · architecture

The builder stack.

A gateway over core builder services, a testing engine, a governed execution engine, planning, and integration into the Knowledge Fabric.

BandLayerComponentsRole
01API gatewayAuthentication · rate limiting · request routing · healthEvery request authenticated and routed
02Core servicesAgent · domain & schema · widget · dataset · utility · chainThe builder services themselves
03Testing & previewSandbox runtime · mock data · debug console · preview renderValidate before publish
04Execution enginePipeline runtime · resource manager · credential vault · state managerIsolated, governed execution
05PlanningHybrid planner · workflow executor · error handler · recovery managerBuild and run from language
06Platform integrationKnowledge Fabric · model services · audit servicesGrounded in the fabric, model-agnostic

The Studio does not only consume the Fabric. It enriches it.

Every governed schema, accepted relationship, published agency, workflow, human decision and execution outcome can cascade back into the Fabric — with provenance, version and authority attached. The enterprise does not merely automate work; its operational knowledge compounds as the work is performed.

Governed schemaAccepted relationshipPublished agencyWorkflow logicHuman decisionExecution outcomeReusable artefact
Drawn and authored

The Fabric begins drawn — discovered automatically from the estate. Through the Studio, people also author: schemas, business rules, approval structures, workflow logic, domain interpretation and reusable agencies. Those authored objects become part of the persistent knowledge state. The architecture is neither purely drawn nor purely authored.

03 · data-driven agents

The unit of execution.

An agent pairs a language model with a structured query plan for schema-bounded processing — not a loose prompt, but a governed, reusable unit with a lifecycle.

Definition

Name, description, instructions, model and system prompt — behaviour stated in plain language.

Query plan

The ordered sources and steps the agent runs: datasets, connectors, sub-agents and scripts.

Placeholders

Configurable slots mapped per user at runtime, so one agent serves many contexts.

Lifecycle

Draft, publish, execute with files, and a full execution history — every run schema-bounded and sandboxed.

01Start a new agent

A user initiates creation. Nothing about the process assumes engineering knowledge.

02Pick or discover a domain

Select an existing business domain, or create one by uploading documents — schemas are extracted automatically from what you upload.

03Review the schema

View the generated schemas and shape them: entities added, modified or removed; attribute types and constraints; relationships defined; the domain assigned.

04Define the agent

Name, description, plain-language instructions, model selection and prompt — then the query plan over datasets, connectors, sub-agents and scripts, with configurable placeholders.

05Draft and test

Save as a draft and test interactively: run with sample files in an isolated sandbox, validate output against the schema, iterate, and apply the draft when satisfied.

06Publish and govern

Move from draft to published, configure runtime placeholder mapping, set scoped access permissions, and make it available to others as reusable supply.

04 · domains & schemas

Your schemas, from your documents.

Upload the documents you already have — specifications, policies, contracts, forms — and the Studio extracts the schemas. The domain model is derived, not declared by hand.

Stage 1

Upload

Provide the documents you already run on.

Stage 2

Analyse

Text, structure and tables are extracted.

Stage 3

Extract

Entities, attributes and relationships discovered.

Stage 4

Generate

Structured schemas created.

Stage 5

Refine

You review and adjust interactively.

Document typeExtraction focus
Data specificationsData models, field definitions
PoliciesWorkflows, rules, roles
Contracts & agreementsParties, terms, obligations
Regulatory documentsRequirements, controls
Forms & templatesInput fields, validations
Schema bindingPurpose
Input schemaValidates incoming data structure
Output schemaEnsures output conforms
Internal schemaControls intermediate transforms
Validation schemaEnforces business rules

Once bound, schemas govern every agent, and agents pin to explicit schema versions.

05 · build without code

Describe it. See the pipeline.

Write the workflow in a sentence; the Studio generates a pipeline you can preview, refine and run — or drag it together from the node palette. Either way, no engineering bench required.

Knowledge & Agentic Studio — text to pipelinevalidated live
You

“Extract client data from the CRM, check it against the sanctions list, and flag high-risk matches.”

Generated pipeline

Input Knowledge (CRM) Tool (sanctions check) If / Else (risk) Output (flag matches) type checking · cycle detection · reachability · permission checks — all validated as you build

The node palette — twelve building blocks

Input — receiveLLM — reasonCode — logicIf / Else — routeClassifier — intentKnowledge — fabric queryHTTP — externalTemplate — formatAggregator — mergeIteration — loopTool — MCP or fabricOutput — return

06 · orchestration

Agencies, not agents.

An agency is a coordinated team of agents serving one purpose. Compose several into a workflow, with human review gates wherever judgement belongs. Execution pauses at the gate until a person approves — then continues. Every step signed, isolated and logged.

Pattern

Sequential with a gate

A → B → [review] → C → D. The straight line, with judgement inserted where it matters.

Pattern

Branching

A splits to B and C, each reviewed, then merged into E.

Pattern

Hierarchical

A delegates to two sub-chains that each run their own steps.

Fig. 01  //  one agency, and where the person sits◆ the gate blocks until approved
ONE AGENCY · ONE PURPOSE · MEMBERS ARE AGENTS, NOT PEOPLETHE AGENCYA drafting agency: six agents, one purpose — a decision-ready response.Retrievereads the estate in placeExtractfields, with confidenceReasonagainst the firm's rulesCheckcompleteness and coverageDraftevery line traced to sourceReviewheld for a personAGENCY6 agentsWHERE THE PERSON SITSExecution pauses at the gate. It does not continue until a person approves.01PrepareThe agency gathers, resolves and reasons.Evidence assembled, not asserted.02Human review gateHeld at the gate. The proposal, the basisfor it, and the approver are recorded.03ActOnly after approval. Signed, isolated,and written to the audit trail.AUTONOMY IS SET PER STEP, NOT PER AGENCY · SEE THE INTERACTIVE DIAL ON THE PLATFORM PAGE
Fig. 02  //  the reusable pattern behind every utility◆ grounded, gated, evidenced
VENDOR CONTROL PLANE reached over PRIVATE LINK (no public path) REGION — [dedicated region · set at deployment] Egress control — allow-list / firewall · default deny AVAILABILITY ZONE A Ingress subnet — private endpoints only[CIDR set at deployment] · no public ingress (P/S) App subnet — data plane computeresolution · engine · gateway · console[CIDR set at deployment] Data subnet — graph storeno egress route · encrypted at rest[CIDR set at deployment] AVAILABILITY ZONE B (HA replica) Ingress subnet — private endpoints only[CIDR set at deployment] App subnet — data plane computehorizontally scaled · same identity model[CIDR set at deployment] Data subnet — graph replicasynchronous / async per RPO[CIDR set at deployment] replicate
Governed multi-agent

Downstream agents can never exceed upstream permissions. Inter-agent messages are signed with a time-to-live. Every human gate blocks until approved, and the approval is recorded with the person who gave it and the basis they had.

Graph of agencies — planned

The next step extends the linear chain into a full directed graph: conditional branching, parallel fan-out and fan-in, cycles with exit conditions and dynamic routing, with a human gate available at any node.

07 · connect anything

Author your own connectors.

Connectivity is MCP First. When a system you built in-house has no built-in connector, you do not wait for the platform — you author one, have it validated, and publish it to the catalogue yourself.

Unified tool catalogue

Built-in nodes, Knowledge Fabric services and connectors sit in one searchable place. Credentials are injected at execution time and never stored in an agent definition.

01

Author

Define the connector type against the MCP contract.

02

Upload

Command, arguments, tools and credentials schema.

03

Validate

Protocol, schema, tool and security checks run before anything is enabled.

04

Activate

Enabled for your tenant on success.

05

Publish

Appears in your catalogue for instance creation by anyone entitled to it.

08 · governance

Governed by construction.

Every agent runs inside an isolated sandbox, calls only the tools it is authorised for, receives credentials only at execution, and writes a tamper-evident trail. Governance is not configured on — it is the default.

No exfiltration

Execution sandbox

gVisor isolation, read-only filesystem, allowlisted network, time and memory limits.

Per use case

Tool authorisation

Read-only is automatic, writes are logged, and administrative actions require multiple parties.

Never in definitions

Credential injection

AES-256, hardware-security-module backed, injected at run and destroyed after.

Before publish

Testing & isolation

Unit, integration, security and performance testing in a separate sandbox.

1–2 year retention

Audit trail

Hash-chained and encrypted, auditor role only, with no delete capability.

Nothing an agency does becomes enterprise truth on its own.

Observed → proposed → validated → approved — and superseded when the evidence changes

What it builds

An encoded workflow is an asset, not an engagement.

The judgement your people apply a hundred times a week is currently held in their heads and lost when they leave. Encoded once in the Studio, it becomes a governed, versioned, reusable object the organisation owns outright — deployable across every team, every client and every jurisdiction, without a vendor in the loop.

Next step

Build the first agency this week.

The people who know the work compose the agencies, from documents they already have. Bring one workflow and the documents behind it.