> ## Content Index > Fetch the complete content index at: https://venkatarayala.com/llms.txt > Use this file to discover other available public pages before exploring further. # The Agent Is No Longer the Architecture: How Salesforce Is Becoming an Agentic Platform - URL: https://venkatarayala.com/agent-no-longer-architecture-salesforce-agentic-platform/ - Published: 2026-09-22T01:04:00.000Z - Updated: 2026-09-23T05:07:35.000Z - Description: Salesforce is moving beyond the individual AI agent toward a composable agentic platform where experience, agency, reasoning, enterprise capabilities, context, and trust operate as distinct architectural responsibilities. - Author: Venkata Rayala - Tags: Agentforce, Salesforce, Agentic AI, Enterprise Architecture, AI Architecture, AIforce, Data 360, Enterprise AI Harness, Headless 360, AI Strategy Dreamforce 2026 introduced a long list of Salesforce AI announcements: **AIforce, Koa, Agentforce Coworker, Headless 360, Data 360, Claudeforce, expanded model choice, and the Enterprise AI Harness.** Individually, each announcement is interesting. Together, they point to something more consequential. For the last few years, enterprise AI conversations have largely centered on the agent: give an agent instructions, provide context, connect actions, select a model, establish guardrails, and put it to work. That model is not disappearing. But Salesforce's emerging direction suggests that the architecture **around the agent** is becoming just as important as the agent itself. The clearest signal may be Salesforce's **Enterprise AI Harness**, which separates the responsibilities AI needs to understand the business, reason, take action, and operate within enterprise controls. That separation matters. Context can be shared. Models can change. Agents can specialize. Business capabilities can be reused. Interfaces can evolve. Security and governance can remain consistent underneath them. The agent is still important. But increasingly, **the agent is becoming one participant in a broader platform architecture, rather than the architecture itself.** > **The agent is no longer the architecture. The architecture is the system that allows intelligence, context, capabilities, and control to work together.** --- ## The Architecture Behind the Agent Before considering individual products, it helps to look at the responsibilities an enterprise agentic architecture must fulfill. From an architecture perspective, I see six distinct layers emerging. ### Intelligent Experience Where people and agents interact with the platform. **AIforce | Coworker | Claude | Gemini Enterprise | Slack | Other Experiences** ### Agency and Orchestration Where intent is understood, work is routed, and specialized agents coordinate. **Agentforce | Specialist Agents | Routing | Coordination** ### Reasoning Models Where intelligence supports interpretation, planning, and reasoning. **Koa | Claude | Gemini | Other Models** ### Actions and Enterprise Capabilities Where reasoning becomes governed enterprise work. **Flow | Apex | APIs | Skills | Workflows** ### Context Where agents obtain the business meaning needed to reason effectively. **Data 360 | Data | Metadata | Semantics | Knowledge | Memory** ### Trusted Platform Foundation Where enterprise boundaries remain enforceable. **Identity | Permissions | Security | Policy | Governance** This is my architectural interpretation, not Salesforce's literal product-stack or runtime diagram. What matters architecturally is the **separation of responsibilities**. > **The interface delivers the experience. The agent provides agency. The model provides reasoning. Capabilities perform the work. Data 360 provides context. The Salesforce platform provides trust and control.** ![](https://storage.ghost.io/c/0b/26/0b260cfa-c7a4-4700-b8c3-a6603829f328/content/images/2026/09/SF-AIPlatform3.png) *An architectural interpretation of the responsibilities surrounding an enterprise agent.* --- ## Coworker and AIforce: Simplifying the Employee Experience The top of this architecture is where complexity begins to disappear from the employee's perspective. **Agentforce Coworker** provides a useful way to understand this shift. Coworker becomes an employee-facing AI teammate while specialized Agentforce agents can operate behind that experience. An employee should not necessarily need to understand: - which specialist agent should handle a request - which reasoning model should be used - where the relevant context resides - which business capability needs to execute - which permissions and policies apply The employee has an intent. The platform handles the complexity behind it. A house provides a useful metaphor. **Coworker is the front door.** Behind that simple entrance sits an architecture of orchestration, specialized agents, reasoning models, business capabilities, context, security, and governance. The employee does not need the floor plan. The architect does. > **The employee sees the front door. The architect must understand the entire house.** AIforce expands this idea beyond a single front door. AIforce brings the Salesforce platform toward intelligent experiences wherever work happens, rather than requiring every interaction to originate from a traditional Salesforce user interface. That creates an important architectural separation: > **Where work happens and where the trusted business capability resides no longer have to be the same place.** ![](https://storage.ghost.io/c/0b/26/0b260cfa-c7a4-4700-b8c3-a6603829f328/content/images/2026/09/image-15.png) *The employee sees a simple experience. Behind it sits an architecture of agency, reasoning, capabilities, context, and enterprise controls.* --- ## Agentforce and Model Choice: Agency Is Not Reasoning Once a request enters the architecture, another separation becomes important. **The agent and the model are not the same thing.** An agent can own a business responsibility while relying on an appropriate reasoning model to supply the intelligence needed for that responsibility. **Koa** makes this distinction especially visible. Koa is Salesforce's CRM reasoning model for Agentforce, built on NVIDIA Nemotron and designed for reasoning through complex, multi-step enterprise workflows. But Koa is only one part of the emerging model story. **Claudeforce makes the composability particularly tangible.** Salesforce can bring its business context, workflows, logic, actions, and governance into Claude. At the same time, Claude can provide reasoning within the Salesforce agentic ecosystem. The architectural significance is bigger than the Salesforce-Anthropic collaboration: **The experience, reasoning model, enterprise capabilities, and underlying platform no longer have to be the same thing.** Google Cloud reinforces the same architectural pattern. **Gemini models can provide reasoning within Agentforce, while Salesforce data and capabilities can also surface within Gemini Enterprise through Salesforce's headless architecture.** This introduces a particularly interesting pattern: **The model can come to Salesforce, or Salesforce capabilities can go to the AI experience.** That is a major architectural shift. The lesson is bigger than **Koa, Claude, or Gemini**. Different workloads may benefit from different reasoning capabilities. The responsibility belongs to the **agent**. The intelligence used to fulfill that responsibility comes from the **reasoning model**. > **The agent owns the job. The model supplies reasoning.** That distinction creates architectural flexibility. If the model becomes the architecture, changing the model risks becoming an architectural redesign. If the model remains a component, model evolution becomes easier to absorb. ![](https://storage.ghost.io/c/0b/26/0b260cfa-c7a4-4700-b8c3-a6603829f328/content/images/2026/09/image-14.png) *Specialized agents own business responsibilities. Reasoning models provide intelligence appropriate to the workload.* --- ## An Open, Multi-Model Agentic Platform This leads to a broader observation. Salesforce's emerging architecture is not simply becoming **multi-model**. It is becoming increasingly open across models, experiences, agents, capabilities, data platforms, and infrastructure. The collaborations surrounding the platform illustrate different parts of this architecture. ### NVIDIA NVIDIA Nemotron provides the foundation on which Salesforce developed Koa, its specialized CRM reasoning model. ### Anthropic Claude brings another reasoning option into the Salesforce ecosystem, while Claudeforce demonstrates that Salesforce capabilities can also extend into Claude. ### Google Cloud Gemini provides another reasoning path, while Salesforce capabilities can extend into Gemini Enterprise. The relationship also extends beyond the model layer into areas such as Data 360 and BigQuery interoperability and Hyperforce on Google Cloud. ### AWS AWS remains another important part of this ecosystem, including cloud infrastructure, Amazon Bedrock, model access, data interoperability, and other Salesforce-AWS integration capabilities. These relationships are not interchangeable. They operate at different architectural boundaries. But together they reveal something important. Salesforce does not need to approach the Agentic Enterprise as one vertically integrated AI stack. Instead: > **Intelligence can come to Salesforce, or Salesforce context and capabilities can go to external intelligence.** That changes the role of the enterprise platform. The strategic value may no longer come from owning every model or every interface. It may come from providing the **trusted context, reusable capabilities, agency, security, policies, and governance that allow different forms of intelligence to participate safely.** --- ## From Applications to Reusable Enterprise Capabilities Reasoning alone does not execute enterprise work. Once an agent determines what should happen, something still needs to perform the governed business action. In Salesforce, many of those capabilities may already exist through mechanisms such as: **Flow, Apex, APIs, Skills, workflows, and established business processes.** This is where traditional enterprise architecture becomes particularly valuable in the agentic era. Instead of embedding business behavior independently inside a Salesforce screen, an integration, and an agent, architects can increasingly think in terms of a **reusable business capability**. The experience requesting the capability may change. The governed business behavior underneath it should not need to be reinvented each time. This is also where **Headless 360** becomes important. Headless architecture makes Salesforce capabilities available beyond the traditional Salesforce user interface. The architectural question therefore begins to change. Instead of asking: > **What screen should expose this functionality?** we can increasingly ask: > **What is the business capability, and which authorized experiences or agents should be allowed to consume it?** Consider something as simple as creating a follow-up. The capability may be invoked from a Salesforce experience today. Tomorrow, an authorized agent or external AI experience might invoke that same capability. The consumer changes. The underlying capability, policy, and governance should remain consistent. > **Build the capability once. Let authorized experiences consume it appropriately.** ![](https://storage.ghost.io/c/0b/26/0b260cfa-c7a4-4700-b8c3-a6603829f328/content/images/2026/09/image-12.png) *Agentic architecture separates reusable business capabilities from the experiences that consume them.* --- ## Data 360: Context Becomes Infrastructure Agents need capabilities through which they can act. But they also need context within which they can reason. This is where **Data 360** becomes important to the architecture. Data 360's role goes beyond simply providing another repository of information. An enterprise agent needs more than records. The agent needs to understand what those records **mean within the business**. I therefore think about Data 360's architectural role primarily as: **Context.** Data provides facts. Metadata provides description and structure. Semantics provides business meaning. Knowledge contributes broader understanding. Memory provides continuity. Together, these contribute to the context within which reasoning becomes useful. > **Agents need more than access to information. They need a governed understanding of what that information means.** This also changes how we should think about metadata. Historically, metadata quality may have been viewed primarily as a platform-management concern. In an agentic architecture, metadata increasingly participates in the context machines use to understand the enterprise. A poorly described environment is no longer simply harder for humans to maintain. It can become harder for AI to interpret correctly. There is a much deeper architecture discussion here around metadata quality, semantics, Salesforce org health, reusable capabilities, knowledge, memory, and enterprise context. That deserves its own discussion. --- ## Trust Must Remain Beneath Reasoning As agents move from answering questions toward performing enterprise work, another architectural boundary becomes critical. **AI reasoning is probabilistic. Enterprise controls often cannot be.** Models can interpret intent. They can reason over information. They can develop plans. They can identify an appropriate capability. But identity, permissions, security policies, approvals, sharing rules, and governance still need to determine what is allowed. This creates a fundamental architecture principle: > **Let AI reason about the work. Don't let AI define the boundaries within which the work is allowed.** The reasoning layer can evolve rapidly. The trust boundary must remain enforceable. That distinction becomes especially important as enterprises introduce multiple models and external AI experiences. The more open the intelligence layer becomes, the more important a consistent trust foundation becomes underneath it. --- ## Cloud Infrastructure Is a Separate Architectural Dimension There is another separation worth making. **Model choice and cloud infrastructure are related, but they are not the same architectural decision.** Hyperforce represents the cloud infrastructure architecture underneath Salesforce, and Salesforce's expansion of Hyperforce across public-cloud environments adds another form of architectural choice. But the infrastructure hosting the Salesforce platform should not be confused with the reasoning model selected for an agentic workload. Conceptually, an architecture may involve: - Salesforce running through Hyperforce - agents operating through Agentforce - business context supplied through Salesforce and Data 360 - enterprise capabilities exposed through workflows, APIs, MCP, and other mechanisms - reasoning supplied by an appropriate supported model These are distinct architectural responsibilities. > **Where the platform runs and how the agent reasons should not be treated as the same decision.** There is an important deeper discussion underneath this distinction. In a multi-cloud, multi-model architecture, architects will increasingly need to understand: - where enterprise data resides - where grounding occurs - what context is supplied to a reasoning model - where inference occurs - what crosses a cloud boundary - how retention policies are enforced - how regional and regulatory requirements apply Those questions matter. But they belong to the physical architecture and trust-boundary discussion rather than the logical architectural model being explored here. For now, the important principle is the separation itself. --- ## Five Principles for Architects The technology and product names will continue to evolve. The architecture principles underneath them are more durable. ### 1\. Separate the Agent from the Architecture Keep capabilities, context, policy, and controls independent from whichever agent currently consumes them. ### 2\. Separate the Agent from the Model Treat reasoning models as components that can evolve independently wherever the enterprise architecture permits. ### 3\. Build Capabilities Before Experiences Create reusable business capabilities, then expose them deliberately through authorized interfaces and agents. ### 4\. Keep Deterministic Controls Beneath Probabilistic Reasoning Reasoning can be adaptive. Identity, authorization, policy, security, and governance must remain enforceable. ### 5\. Treat Context as Architecture Data alone is insufficient. Metadata, semantics, knowledge, memory, and business meaning increasingly shape whether AI reasoning is relevant to the enterprise. --- ## The Durable Advantage Is Architectural Optionality Models will change. Agents will specialize. Interfaces will multiply. Cloud and AI ecosystems will continue to evolve. The experience employees use today may eventually become one of many. The durable enterprise assets are different. They are the **trusted context, reusable capabilities, identity, permissions, workflows, policies, security, and governance** underneath those experiences. Those are assets enterprises have spent years building. The agentic era does not make them less relevant. It makes their value more visible. Once experience, agency, reasoning, action, context, control, and infrastructure can evolve with greater independence, enterprises gain something more durable than access to the latest model. They gain **architectural optionality**. A new model does not necessarily require rebuilding the business capability. A new agent does not necessarily require recreating context. A new interface does not necessarily require duplicating business logic. A new cloud infrastructure option does not necessarily have to redefine the reasoning architecture. And stronger intelligence does not require weakening enterprise controls. That, to me, is the architectural significance of the direction Salesforce revealed around Dreamforce 2026. The agent still matters. But the architecture surrounding the agent increasingly determines whether that intelligence can become dependable enterprise execution. > **The agent is becoming one participant in the architecture, not the architecture itself.** ## References - [Salesforce Introduces the Trusted Enterprise AI Harness](https://www.salesforce.com/news/stories/enterprise-ai-harness/?ref=venkatarayala.com) - [Salesforce Unveils AIforce, Bringing the Full Power of Its Platform to Any Interface](https://www.salesforce.com/news/stories/aiforce-announcement/?ref=venkatarayala.com) - [Announcing Koa: Salesforce’s First CRM Reasoning Model, Built on NVIDIA Nemotron](https://investor.salesforce.com/news/news-details/2026/Announcing-Koa-Salesforces-First-CRM-Reasoning-Model-Built-on-NVIDIA-Nemotron/default.aspx?ref=venkatarayala.com) - [Salesforce and Anthropic Announce Claudeforce](https://www.salesforce.com/news/press-releases/2026/08/26/salesforce-and-anthropic-announce-claudeforce/?ref=venkatarayala.com) - [Salesforce and Google Cloud Unify Infrastructure and Agents](https://www.salesforce.com/news/stories/salesforce-google-cloud-unify-infrastructure-and-agents/?ref=venkatarayala.com) - [Introducing Salesforce Headless 360](https://www.salesforce.com/news/stories/salesforce-headless-360-announcement/?ref=venkatarayala.com) - [AWS and Salesforce Expand Collaboration for Enterprise AI](https://www.salesforce.com/news/stories/aws-salesforce-enterprise-ai-expansion/?ref=venkatarayala.com)