Allied EraThree Ideas, One Direction
If you follow Salesforce news, three terms keep coming up: Agentforce, headless agents, and MCP. They can sound like three separate products, but they all point in the same direction:
Salesforce is becoming a platform that AI can use directly — not only people clicking through screens.
- Agentforce lets you build AI agents inside Salesforce.
- Headless agents let other systems use those agents without a chat window.
- MCP lets AI assistants outside Salesforce (like Claude) safely use Salesforce data and tools.
This guide explains each one in plain language, shows how they fit together, and helps you decide which one your business needs.
Part 1: Agentforce — AI Agents Built Inside Salesforce
Agentforce is Salesforce's platform for building AI agents that understand requests, use your Salesforce data, and take action.
The building blocks
| Building block | What it means | Example |
|---|---|---|
| Topics | The jobs the agent handles | "Order status", "Returns", "Room booking" |
| Instructions | Plain-language rules for each job | "Always verify the customer's email before sharing order details." |
| Actions | What the agent is allowed to do — built with Flows, Apex, and prompt templates | "Get order status", "Create case", "Book room" |
| Reasoning engine | Decides which topic and actions fit each request, and in what order | Picks "Order status" → runs "Get order status" |
| Trust layer | Salesforce's security layer around the AI (data masking, toxicity checks, audit trail) | Hides sensitive values from the language model |
Where Agentforce agents usually live
Most people meet Agentforce through a conversation: a chat window on a website, a messaging app like WhatsApp, a customer portal, Slack, or the assistant panel inside Salesforce. That is the "with a head" version — there is a user interface where a person talks to the agent.
But that is only half the story.
Part 2: Headless Agents — Agents Without a Chat Window
A headless agent is an Agentforce agent that runs without a user interface. Instead of a person typing in a chat box, a system calls the agent through code, and the agent's answer goes back to that system.
Think of it like this: a normal agent is a receptionist talking to visitors. A headless agent is a back-office specialist who receives work through an internal system and sends results back — no one sees them, but the work gets done.
Three ways to call a headless agent
| Method | Who uses it | How it works |
|---|---|---|
| Agent API | Developers connecting external systems (websites, mobile apps, ERPs, other AI agents) | The system authenticates with Salesforce, starts a session, sends messages, receives responses, and ends the session. |
| Flow (AI Agent action) | Admins | A Flow step sends a message to an agent and uses the reply in later steps. |
| Apex | Developers | Apex calls the agent through an invocable action, for example from a trigger or scheduled job. |
The Agent API supports two styles of response:
- Synchronous — the caller waits for the full answer. Simple, and good for background jobs.
- Streaming — the answer arrives in pieces (using server-sent events), including progress messages. Better for custom apps where a user is watching.
What can headless agents do?
- Your own app, Salesforce's brain: a mobile app or website with its own design uses an Agentforce agent behind the scenes.
- Background decisions: instead of a Flow with dozens of branches, one step asks an agent to classify a request and decide the next step.
- Multi-agent processes: for example, a loan application where one agent talks to the customer and others check documents, score risk, and recommend a decision.
- Data quality checks: a scheduled job asks an agent to review new records and flag problems before users notice them.
- Connecting other platforms: an external system sends one request to an agent instead of making many separate API calls.
Things to watch
- AI decisions are not 100% predictable. Keep rules that must never change (legal, financial) in normal Flows or Apex, and use agents where judgment helps.
- With no human watching the chat, logging and testing matter more. Review what the agent decided, especially in the first weeks.
- Agent API calls need proper authentication set up by an admin, and each call uses agent capacity.
Part 3: MCP — Letting Outside AI Work With Salesforce
MCP (Model Context Protocol) is an open standard for connecting AI assistants to tools and data. It works like a universal plug: one standard way for any compatible AI app to use a system's capabilities.
Headless agents and MCP are mirror images:
- With a headless agent, an outside system asks Salesforce's AI to do the work.
- With MCP, an outside AI (such as Claude) uses Salesforce's data and actions directly.
Salesforce Hosted MCP Servers
Salesforce runs MCP servers for you — there is nothing to host. They became generally available in April 2026. An admin activates them in Setup, and users connect from an MCP-compatible AI app by signing in with their Salesforce account.
- Standard servers cover common needs, such as working with object records (for example, the
sobject-allserver), plus servers for Data 360, Tableau, and MuleSoft. - Custom servers let you publish your own business logic as AI tools — built from invocable Apex actions, autolaunched Flows, Apex REST endpoints,
@AuraEnabledmethods, and prompt templates. This is how you make sure the AI follows your rules, not just raw data access.
The Headless 360 MCP Server (Beta)
In July 2026, Salesforce released the Headless 360 MCP Server in beta. Instead of exposing thousands of separate tools, it exposes four simple ones and a growing library of skills (step-by-step playbooks for tasks):
| Tool | What it does |
|---|---|
| Discover | Searches for the right skills and APIs for a goal, in plain language |
| Describe | Explains a skill: which APIs, which inputs, and in what order |
| Dispatch | Runs the skill, after checking the user's access |
| Dispatch Read Only | Runs look-up actions that never change anything |
It launched with around 100 skills, mostly for admin and setup work — such as managing users, Apex triggers, platform events and Change Data Capture, and named credentials — with more cloud-specific skills planned.
Example: an admin asks an AI assistant, "Send Account changes to our AWS system." The assistant uses Discover to find the right skills, Describe to learn the order (for example, create a named credential first), and Dispatch to carry it out — with the admin able to approve each step.
Is MCP secure?
- Every action runs as the signed-in user, through an External Client App that an admin controls.
- Normal Salesforce security still applies: object and field permissions, sharing rules, profiles, and permission sets. If a user can't do it in Salesforce, the AI can't either.
- Actions appear in the audit trail under the user's name.
- Most AI apps can ask for approval before each tool runs.
Salesforce also recommends testing configuration changes in a sandbox before production — especially because admins and developers have broad permissions.
Part 4: Headless 360 — The Big Picture
Headless 360 is Salesforce's name for this overall strategy: turning the platform into reusable business capabilities that AI agents can discover and use securely, while reusing the identity, permissions, metadata, and business logic you already have in Salesforce.
By August 2026, Headless 360 covered several MCP servers (Headless 360, Data 360, and Marketing Engagement), more than 100 reusable agent skills, a headless experience layer for building web, mobile, and conversational apps, and coverage across sales, service, marketing, commerce, MuleSoft, Informatica, and Tableau.
In short: your Salesforce investment becomes usable by any approved AI, not just by people using Salesforce screens.
How They Fit Together
| Agentforce (with UI) | Headless agent | MCP | |
|---|---|---|---|
| Where the AI runs | Inside Salesforce | Inside Salesforce | Outside Salesforce (e.g., Claude) |
| Who talks to it | A person, in a chat | A system, through code | A person, in their own AI app |
| Best for | Customer service, employee assistants | Custom apps, automation, multi-agent processes | Admins, developers, analysts, power users |
| Who builds it | Admins and developers in Agentforce Builder | Agentforce Builder + developers for the integration | Admin activates servers; developers add custom tools |
They also combine well. For example, a company can run an Agentforce chat agent for customers, call the same agent headlessly from its mobile app, and let its admins manage Salesforce setup through Claude using the Headless 360 MCP Server — all on the same data, rules, and permissions.
Which One Do You Need?
- "I want customers to get answers 24/7 on our website or WhatsApp." → Agentforce with a chat channel.
- "I want AI inside our own app or automated in the background." → A headless agent through the Agent API, Flow, or Apex.
- "I want my team to ask Claude questions about Salesforce data, or speed up admin work." → Salesforce Hosted MCP Servers.
- "I want AI tools to follow our exact business rules." → Custom MCP tools built from your Flows and Apex.
How to Start Safely
- Pick one clear job — for example, order status questions or a weekly data-quality check.
- Work in a sandbox first, with test data.
- Review permissions — the AI inherits them, so make sure users only have the access they need.
- Start read-only, then add actions that change data one at a time.
- Test and monitor — check real conversations and decisions during the first weeks.
Quick Glossary
- Agent: an AI that can understand a request and take actions, not only answer.
- Headless: no user interface; used by other systems through code.
- Agent API: Salesforce's API for talking to Agentforce agents from outside.
- MCP: Model Context Protocol, an open standard for connecting AI apps to tools and data.
- MCP server: the connector that exposes a system's tools to AI apps.
- Skill: a packaged, step-by-step playbook an AI can follow to complete a task.
- Headless 360: Salesforce's strategy for making all its capabilities available to AI agents.
Conclusion
Agentforce, headless agents, and MCP are three doors into the same house. Agentforce lets you build agents inside Salesforce, headless agents let any system use them, and MCP lets outside AI tools work with Salesforce safely. Together, they turn Salesforce from a place where people enter data into a platform that AI can act on — with your security rules still in charge.
Want help choosing the right approach for your business, or building your first agent? Contact Allied Era.
Sources: Salesforce Developers Blog (Agent API, Hosted MCP Servers, Headless 360 MCP Server beta) and Salesforce Headless 360 announcements, 2025–2026. Features in beta may change.
Comments (0)
Loading comments...