Architecture comparison
Proposed patent architecture against publicly documented architectures
Every company is mapped against the same eleven-layer stack. Entries describe only what is identified in reviewed public material for that company; blanks are recorded expressly rather than left implied.
| Proposed architecture layer | Adobe | Microsoft | Meta | AWS | Salesforce | NVIDIA | |
|---|---|---|---|---|---|---|---|
| 01 External Digital Platforms / Enterprise Systems Independent social, streaming, commerce, CRM, cloud and AI environments connected as sources and execution targets. | Experience Cloud applications, Creative Cloud, commerce and enterprise data sources connected through Adobe APIs. | Microsoft 365, Dynamics 365, Azure services and hundreds of documented connectors to external systems. | Facebook, Instagram, Threads and WhatsApp surfaces; cross-app activity where publicly documented. | Google Cloud, Workspace, Search, YouTube, advertising systems and third-party systems reachable through A2A and MCP. | Any external system reachable through AgentCore Gateway, API integrations and event sources. | CRM, Commerce, Marketing, Slack, Tableau, MuleSoft-connected enterprise systems and external APIs. | Enterprise and industrial systems, sensor and simulation data sources. |
| 02 Intelligent AI Orchestration Layer Central coordination of models, agents, data, workflows and external actions across otherwise unconnected environments. | GenStudio is publicly positioned as a connective layer across the content supply chain; orchestration is described within the Adobe ecosystem rather than across arbitrary independent third-party platforms. | Generative orchestration in Copilot Studio and the Foundry Agent Service select tools, knowledge and sub-agents at run time. | Not identified in reviewed public material as a generalised cross-platform orchestration layer; coordination is internal to Meta surfaces. | Vertex AI Agent Engine and ADK provide orchestration primitives; a creator-ecosystem orchestration layer is not identified in reviewed public material. | AgentCore Runtime and Step Functions provide orchestration primitives; a domain orchestration layer is not identified in reviewed public material. | Agentforce orchestrates agents, actions and data access under enterprise guardrails. | Agent toolkit and inference orchestration across microservices. |
| 03 Autonomous AI Agents Creator, Viewer, Commercial, Sponsorship, Marketplace, Rights, Distribution, Enterprise, Brand, Presentation and Orchestration agents. | Adobe publicly describes AI agents across the content supply chain, including brand and production-oriented agents. | Agents and connected agents, including delegation to specialised agents. | Creator Assistant and Meta AI assistants are publicly described; a family of commercial, rights and marketplace agents is not identified in reviewed public material. | Agents built with ADK and deployed on Agent Engine; interoperating agents via A2A. | Bedrock agents and multi-agent collaboration with a supervising agent. | Service, sales, marketing and commerce agents; custom agents authored on the platform. | Multi-agent systems built on NVIDIA toolkits and models. |
| 04 Digital Twins Behavioural models of creators, viewers, communities, commercial relationships, enterprises and digital content assets. | Not identified in reviewed public material as behavioural Digital Twins; unified customer profiles in Adobe Experience Platform are the nearest documented construct. | Not identified in reviewed public material as behavioural Digital Twins. | Not identified in reviewed public material as Digital Twins; user and creator models exist within recommendation systems. | Not identified in reviewed public material as behavioural Digital Twins. | Not identified in reviewed public material as Digital Twins in the behavioural sense. | Data Cloud unified profile is the nearest documented construct; behavioural Digital Twins are not identified in reviewed public material in those terms. | Omniverse Digital Twins — physical/industrial simulation, a different implementation and objective from behavioural twins. |
| 05 Persistent Behavioural Memory Durable retention of interactions, decisions, outcomes, workflow history and commercial activity across sessions and platforms. | Profile-level persistence and experience event history documented in Adobe Experience Platform; persistent agent memory across sessions not identified in reviewed public material in the same terms as Copilot Studio or AgentCore. | Documented agent memory retaining context across sessions. | Long-horizon behavioural modelling is publicly described for recommendation; assistant goal-learning over time is publicly described for Creator Assistant. | Agent Engine Memory Bank documents long-term memory personalising subsequent sessions. | AgentCore Memory documents short-term and long-term memory across sessions. | Unified profile with harmonised historical engagement and transaction data. | Shared state and memory in agent toolkits; persistent behavioural memory of a person is not identified in reviewed public material. |
| 06 Semantic Knowledge Graph Relationships between creators, viewers, content, communities, brands, rights, agreements, systems and behavioural events. | Adobe Experience Platform data model and identity graph; a semantic knowledge graph spanning creators, rights and agreements is not identified in reviewed public material. | Microsoft Graph and enterprise knowledge sources; a semantic graph across creators, rights and commercial agreements is not identified in reviewed public material. | Entity and interest graph technologies are publicly described; a rights and commercial-agreement graph is not identified in reviewed public material. | Knowledge Graph and entity technologies are publicly documented; a creator-rights-commercial graph is not identified in reviewed public material. | Bedrock Knowledge Bases including graph-backed retrieval. | Data model and relationships across accounts, contacts, products and interactions; a creator-rights graph is not identified in reviewed public material. | Scene graphs and structured simulation state; a semantic creator/rights graph is not identified in reviewed public material. |
| 07 Predictive / Commercial / Audience / Content / Rights Intelligence Forecasting and evaluation layers operating on the twins, memory and graph. | Customer, campaign and content-performance insight and reporting. | Analytics through Fabric and Dynamics; predictive creator/audience intelligence not identified in reviewed public material. | Extremely strong publicly documented engagement and content-performance prediction. | YouTube analytics and advertising prediction are publicly documented; Vertex provides general prediction infrastructure. | General ML and forecasting services; creator/audience prediction not identified in reviewed public material. | Predictive customer intelligence, propensity and next-best-action. | Simulation-based prediction and optimisation of physical processes. |
| 08 Autonomous or Assisted Workflow Decisions Decisions taken autonomously or presented for human approval, with explanation of the underlying reasoning. | Workflow automation and approval within GenStudio and Workfront. | Autonomous triggers and human-in-the-loop approval documented. | Recommendations presented to creators; autonomous cross-platform workflow execution not identified in reviewed public material. | Agentic workflows with tool use; autonomous action documented in agent frameworks. | Agent decisions with tool invocation; human approval patterns documented. | Autonomous and assisted actions with human escalation documented. | Autonomous control decisions in simulated and physical systems. |
| 09 Execution Across External Platforms Publishing, transformation, distribution, notification, commercial and rights actions performed on connected systems. | Activation and delivery across owned channels, advertising and connected destinations. | Execution through connectors, tools and API actions. | Publishing and distribution across Meta surfaces. | Execution via tools, APIs and A2A partner agents. | Execution through gateways, tools and APIs. | Execution through platform actions, flows and MuleSoft-integrated systems. | Execution against industrial systems and robots rather than consumer platforms. |
| 10 Observed Outcomes Measured engagement, commercial, audience, rights and workflow results returned into the architecture. | Reporting and insights on delivered content performance. | Telemetry and analytics on agent runs. | Insights and analytics returned to creators. | Analytics, YouTube Studio insight and advertising measurement. | Observability and evaluation tooling documented. | Campaign, service and commerce outcome measurement. | Sensor and simulation outcome measurement. |
| 11 Continual Learning and Model Evolution Outcomes update behavioural memory and Digital Twins, modifying subsequent orchestration decisions. | Publicly described as insight feeding back into planning and creation; the closest documented feedback loop of any company reviewed at the content-lifecycle level. | Evaluation and continuous improvement tooling documented; automatic write-back of measured outcomes into a behavioural twin is not identified in reviewed public material. | Continual model training from observed engagement is publicly described at platform level. | Memory Bank updates from session outcomes; a full commercial-outcome-to-twin cycle is not identified in reviewed public material. | Evaluation-driven improvement documented; automatic outcome write-back into a behavioural twin is not identified in reviewed public material. | Optimisation from campaign outcomes publicly described; write-back into a per-person behavioural twin is not identified in reviewed public material in those terms. | Twin updated from observed outcomes — structurally comparable feedback loop in a different domain. |
Divergence analysis
Where the layers correspond and where they diverge
Layers with strong public correspondence
- Orchestration layer. Publicly documented by Microsoft, Google, AWS, Salesforce and Adobe as agent runtimes, generative orchestration and content supply chains.
- Agent layer. Specialised agents, delegation and multi-agent coordination are documented across five of the seven companies.
- Memory layer. Persistent cross-session agent memory is a productised primitive at Microsoft, Google and AWS.
- Execution layer. Connector, gateway and API execution against external systems is mature everywhere.
Layers where divergence is identified
- Digital Twin layer. Behavioural twins of creators, viewers, communities and content are not identified in reviewed public material. The nearest constructs are unified customer profiles (Adobe, Salesforce) and physical simulation twins (NVIDIA) — a different implementation and objective.
- Principal alignment. Reviewed architectures act for the platform or the enterprise. The proposed architecture's agents act for the individual creator or viewer.
- Cross-platform scope. Reviewed creator intelligence is confined to the operator's own surfaces.
- Outcome write-back. Per-principal reconciliation of prediction against measured outcome is not identified in reviewed public material.
Critical section
Continual-learning feedback architecture
Observe
Store
Model
Predict
Decide
Act
Measure Outcome
Update Behavioural Memory / Digital Twin
Improve Next Decision
Adobe publicly demonstrates this loop at the level of an enterprise content supply chain; Meta demonstrates it at the level of platform-wide model training; NVIDIA demonstrates it at the level of a physical simulation twin. Closure of the loop at the level of an individual creator or viewer, using a behavioural twin updated from commercial outcomes measured on platforms the operator does not control, is not identified in reviewed public material. Marked: Patent Position Analysis / Technical Interpretation.