The Rise of AI Agents: Salesforce Introduces Agentforce

With Agentforce, Salesforce is pioneering the future of AI-driven workflows, enabling seamless integration across its ecosystem of applications and services

I’m old enough to remember the launch of Salesforce’s AppExchange in 2006. It was a pivotal moment in the evolution of software as a service, ushering in the age of open SaaS platforms, and a truly seminal moment in the evolution of Salesforce itself. Appexchange redefined what it meant to be a CRM and what customers could do with their CRM data. By making Salesforce data interoperable with partner applications through extensible, open APIs, AppExchange created an array of new operating leverage points for its customers through their own data.    

Enter AgentForce

Salesforce’s recent announcement of Agentforce feels like a similarly transformative moment in the company’s transition to an AI-first platform. With Agentforce, Salesforce is creating new operating leverage for customers by extending Salesforce data and application integrations to drive autonomous AI-driven agentic workflows across the Salesforce platform and partner ecosystem. Agentforce is 2006’s AppExchange, now fortified with a hyper-intelligent borg army.

AgentForce puts Data into Action using the Salesforce Atlas Reasoning Engine. Image Credit: Salesforce

In 2007, when I was a young marketing manager, I connected my startup’s Google Adwords to Salesforce using AppExchange so we could see which paid search keywords were driving inbound leads into our sales funnel. It was a light integration technically, but a big deal for a startup that needed to squeeze maximum efficiency from our marketing budget. Each month, we handed the insights from this reporting integration to a search marketing agency that would go back into Google Ads and optimize our keyword bidding strategy for us.

In the Intelligence Age, a better approach might be to have an AI agent in Salesforce go interact with an AI agent in Google Ads to perform that value-based bid optimization automatically without the need for any human intervention. This is fundamentally the vision for how Agentforce will work with the thousands of applications that interface with the Salesforce platform today. You identify a workflow or repetitive task that you no longer want to deal with manually, you setup AI agents to take over these tasks, and as a fail-safe, you set guardrails to deflect the task to humans when the agents are confounded by anything you didn’t anticipate when building the agent.

“Agentforce’s limitless digital workforce of AI agents can analyze data, make decisions, and take action on tasks like answering customer service inquiries, qualifying sales leads, and optimizing marketing campaigns. With Agentforce, any organization can easily build, customize, and deploy their own agents for any use case across any industry.”

— Salesforce Agentforce Launch Press Release

Agentic Network Effects

Marc Benioff could not have possibly foreseen generative AI back in 2006, but he did understand the power of network effects, and boy is he about to get a huge network effects dividend for building an app ecosystem around Salesforce. According to Salesforce, the average enterprise has over 1000 applications deployed, which is a very complex technology terrain for AI agents to navigate. Salesforce Agents, empowered to work across the Salesforce partner ecosystem, should be capable of becoming some of the very best agents.

For AI agents to be effective, they need other agents and related tools to complete actions. The more agents and tools they have access to, and the more well-versed they are in the use of those tools, the better they can be at, well, being good agents. Imagine if a handyman came to your house and all they brought was a wrench, there would only be a limited set of problems they could help you with. If they brought a toolbox, there are a larger number of problems they could help you fix. If they brought a large tool truck and all the manuals for your home systems and appliances, they could probably help you with just about any problem you might have in your home… plumbing, electrical, lighting, heating, or whatever else you need.

With Agentforce, Salesforce is bringing the AI agent tool Semi Truck and then some. As Vanilla Ice might say, “Yo, If you gotta problem, AgentForce will solve it…”

One Billion Agents

Salesforce wants to build over a billion agents in the first year of Agentforce’s release. That’s what I like to call a BHAAIG, or Big Hairy Audacious AI Goal. The company’s launch partners include Amazon Web Services, Appiphony, Asymbl, Box, Certinia, Copado, Coupa, Docusign, GoMeddo, Google Cloud, Honeywell, IBM, Korn Ferry, Moody’s, OpenText, Sprout Social, TerraSky, Workday, Zoom, with many more likely to fast follow.

For all this to work, Salesforce must get a few important technical things right. Integrating intelligent agents isn’t like integrating APIs. Moving data from one application to another is a relatively predictable affair. Getting AI agents to reliably follow prompts, select the right tools, utilize those tools correctly, pass information from those tools to other agents, and then know when to complete a task still feels more like prompt art than science at this stage of the AI agent game.

Moreover, getting consistent output from agents, even when repeating the same task, using the same prompts, tools, and agent design, isn’t always guaranteed. If you’ve ever asked ChatGPT the same question twice and gotten different answers, you know what I mean.

DreamForce Demo: How Agentforce Works

All that caution aside, Salesforce appears to have done a very nice job of putting guardrails around what agents can do and how they behave. It’s even included some nifty AI tools for pressure-testing AI agents before deployment. Yes, AI testing AI. That’s the world we live in. Here’s a quick step by step of how it all works from the Dreamforce AgentForce Demo:

Step 1 – Instructions: Tell your agents what they can and cannot do.

To begin building an agent, you can select an out-of-the-box agent or a custom created agent. Agentforce includes a number of out-of-the-box agents for different tasks including a Service Agent for CSR tasks, a Sales Development agent that can qualify leads, a Personal Shopper agent that can help visitors on your website, and a Campaign Optimizer agent for optimizing your marketing campaigns.

To setup an agent, you start by providing instructions to Topics assigned to an agent. Topics are jobs that can be performed by agents. In the case of a Customer Service Agent, Topics might include things like handling FAQs, responding to customer account inquiries, or managing an order. Every topic has two elements: instructions and actions. Within each topic you can provide instructions, these are prompts telling the agent what types of user queries they are designed to respond to and how they should respond.

Instruction builder for Salesforce Agents. Image Credit: Salesforce

In the Agentforce keynote, Salesforce shared an example of an AI customer service agent. The instructions and actions provided to the demo agent included confirming the customer’s identity, confirming order details with the customer, confirming changes to an order and other common CSR-related tasks.

Instructions also provide an opportunity to tell agents what they cannot do. In this simple example, the demo agent was provided an instruction to deflect to a human agent in situations where the customer wanted to change their billing information.

Negative Instructions: Giving agents guardrails. Image Credit: Salesforce

Step 2. Actions: Give agents the tools to act on your instructions

Once you’ve defined instructions for your agent, you need to define the actions they’re allowed to take to fulfill requests. In Agentforce, actions allow agents to act on their instructions. Actions are analogous to function-calling (also sometimes referred to as tool-calling) in the LLM world. This is a clever re-brand of the term by Salesforce marketing, perhaps done to simplify the jargon a bit for non-technical users.

LLMs like Chat GPT4 have just recently become smart enough to interpret the context of a prompt and call available tools or APIs to complete requests. This is the foundation of Agentic AI. Without function-calling, there are no agents.

Assigning Actions to Agents in Agentforce. Image Credit: Salesforce

The most remarkable part of Agentforce is the sheer breadth of actions available out of the box. Agentforce actions are built on existing Salesforce platform tools like Flows, Apex, Mulesoft, and Einstein prompt templates. For customers that have made heavy investments in custom Salesforce implementations, this means it will be an easy transition to Agentforce. They can use their existing workflows, integrations and business logic in Salesforce to power their agents. This is a huge competitive advantage for Salesforce and a big value-add for its customers. You’re certainly not limited to your existing assets; you can build new custom actions for agents on the fly using UX-based tools without writing any code.

Step 3. Test your agents

I know what you’re thinking. What if these agents say crazy things to my customers? How can I be sure they won’t hallucinate. That’s becoming a very last year kind of problem. Salesforce calls Agentforce’s agents “Low-Hallucination” agents. LLMs have gotten a lot smarter and the guardrails a lot tighter now with context-aware, RAG based architectures that can put bounds on what agents do and say. 

Agentforce takes things a step further with native tools for previewing and pressure-testing new agents. Once you’ve completed an agent design, you can easily test it in the Preview pane and see how it responds to requests. In the preview pane you’ll see how the agent responds to requests and what actions it takes as it reasons its way to a final output.

Agent Preview in Agentforce. Image Credit: Salesforce

My favorite feature in the demo was Batch Testing. Rather than manually go through dozens or hundreds of possible customer interactions to pressure test an agent design, Salesforce has created a very useful and fast way of auto-generating customer “utterances” or possible customer interaction scenarios that an agent might encounter.

This allows for nearly instantaneous batch testing across hundreds of interaction scenarios for a new agent. The “Testing Center” tools automatically test the agent response and provide a pass/fail score for each utterance. This is an elegant way to answer enterprise concerns around how to build confidence in, and get controlled output from, intelligent agents. Well done Salesforce.

Batch Testing in Agentforce Testing Center. Image Credit: Salesforce

Conclusion

So that’s Agentforce, or at least what we know thus far.

What we didn’t see in the demo was multi-agent or sequential agent flows across partner applications and platforms. The state of the art in agentic AI right now is to have multiple agents working autonomously to solve a task. With such a large ecosystem of partners, Agentforce must be able to handle multi-agent, cross-platform flows. How reliably Agentforce tooling handles the supervision of these more complex agent scenarios will be the acid-test for customers adopting Agentforce. At launch, only 20 partner agents and actions were available, so customers and Salesforce will most likely be learning together as they move up the maturity curve in agentic AI.

For now, though, what they’ve shared on day one of the launch is pretty awesome. It’s a positively great time to be a Salesforce customer.