Salesforce Agentic Orchestration and Flow

Innovation in agentic orchestration is accelerating rapidly, and our automation product teams have been hard at work improving Flow to make it easier than ever to bring agents and human-centered work together. With the Summer ’25 release, we’ve introduced powerful new capabilities that help you embed Agentforce agents into your automations and extract structured data from agent responses. I also have an update on how we’re tapping into MCP to make Flow and Action Platform capabilities available to the full power of all actions —connect any agent technology to the full power of Salesforce actions.

Add agents to automation with Flow & Agentforce

As of Summer ‘25 release, you can now easily insert Agentforce agents into most flows. 

This video from Liz Awad, one of our product managers, demonstrates a flow that features two different embedded Agentforce agents. One of the great strengths of Flow is its ability orchestrate different agents and combine them with business logic that guarantees certainty and safety (often referred to as deterministic processing).

How It’s Done 

Activate your Agentforce Agent in Setup. Then go to a flow that supports the Action element, such as a Screen Flow, an Autolaunched Flow or a Record-Triggered Flow. Add the Action element, and in its property editor, click on the AI Agent Actions category, and you’ll see an action for each of your activated Agentforce agents.

If you have MuleSoft for Flow : Integration installed, you’ll also be able to find actions for sending prompts to OpenAI and Anthropic.

Easily extract information from Agent responses with Structured Outputs

One of the problems we have had to solve in the last year involves the handling of the text output that Agents return. That text is unstructured, just like this post. But suppose you want to use information that the Agent returned in subsequent parts of your flow? You need a way to accomplish several things:

  1. You need to instruct the agent on which specific data you want to get back. For example you might want to tell it to return a temperature value labelled ‘Temp’ and formatted in celsius degrees.
  2. You need to be able to surface those specific outputs from the agents as resources that can be mapped to the inputs of ‘downstream’ actions. For example, you might want to map the temperature value into an action that calls out to a legacy system that’s very particular about formatting. 

Our Action Platform team worked with our Agentic Orchestration program to solve this by enhancing all of the Agentforce actions to use Structured Outputs, a simple mechanism that lets the builder, at Design Time, identify specific named values that they want to use. These values are passed to the agent as instructions and the agent returns them in a format that the flow can automatically extract. The values also surface as available Design-Time resources that can be mapped to downstream inputs. 

How It’s Done

From the property editor use the Configure Structured Outputs section to define the outputs you want to get back from the agent:

Coming Soon: MCP Server Support for All Invocable Actions

There are now more than 1000 standard actions made available by different Salesforce products, and each month more than 350,000 different custom actions (written in Apex) get used. Most of these are available today in Agentforce, but what if you want to use a different agent technology but still access all of your Salesforce actions and capabilities? 

It will soon be possible to define MCP Servers on the Salesforce Platform. MCP is a new technology that, among other things, standardizes the way that actions can be tapped into by agents. When you create an MCP server, you will be able to add any Salesforce action to it, immediately providing a way to expose actions to your agent of choice. 

Summary: Orchestrate Agents Today with Flow and Actions

You can use generally available Flow and Action features today to orchestrate rich interactions that mix agents and humans. This allows the immediate combination of the reasoning power of agents and the certainty of process automation, which provides a way to insert agentic reasoning into all manner of business processes while maintaining control and safety. I encourage you to try our latest features out!