Home Artificial Intelligence Microdosing on Low-Hallucinogenic AI – Unite.AI

Microdosing on Low-Hallucinogenic AI – Unite.AI

by admin
mm

The Agentforce is here. Salesforce wrapped another edition of its annual Dreamforce conference this September. Joining the swarms of attendees — and the swarms of Waymos shuttling them around an extra-cleaned San Francisco — we each now have a swarm of agents at our fingertips to transform work, neatly controlled within the Salesforce ecosystem. While Dreamforce is always a spectacle for its marketing-honed pronouncements of the future, this year provided an unexpectedly compelling vision of how AI-based agents are about to revolutionize the workplace and customer experience.

Let’s temper our expectations for just a bit longer. Benioff pondered at his keynote, “Why would our agents be so low-hallucinogenic?” Yes, they have the data, the metadata, the workflows and a vast array of services to connect into; and so long as your systems only live within Salesforce, it sounds pretty ideal. Salesforce may or may not have invented prompt engineering, a claim Benioff also made in the keynote evoking perhaps the “Austin Powers” Dr. Evil monologue about his father inventing the question mark. But can Salesforce meet the Agentforce vision? If they do, it’s going to be a big deal for how work gets done.

Let’s be real though: our systems and data don’t all live within Salesforce. If the future of work is defined by groups of agents working together, how far can walled gardens and closed ecosystems really get us in delivering outcomes across our businesses? Certainly Apple and Microsoft and Amazon and a host of others want to wrap their arms around the massive agent opportunity in front of us. But as each wave of technical advancement has brought forward different flavors of closed vs. open debates, we’re ultimately going to need a standard for agents working together with one another across boundaries. Otherwise only parts of your business will meet this opportunity.

As we often do when faced with the open/closed conundrum, let us look to the open web as a way forward. Just like apps on your phone need a web view to enable an infinite array of mobile app outcomes, the same will be needed in the upcoming multi-agent frontier. Tools like Slack provide UI frameworks like Block Kit that can power the user interface for a simple agent interaction, but it’s not cut out to handle the depth of modern user experiences. Take Clockwise Prism as an example. We’ve built a next-level scheduling agent for finding time for a meeting even if there isn’t any current “whitespace” on the calendar tomorrow. When hooking it up to other agents to land that impossible meeting with your hottest sales prospects, you’ll need a way to either confirm or explore a myriad of sophisticated and powerful scheduling options. Providing a web view for doing this is the clear path forward.

Throughout his keynote, Benioff repeated the mantra that you don’t want DIY agents within your business. And he’s right. Enterprises want controlled and simplified workflows that deliver repeatable value. And yet they don’t want to be stuck in a silo. This is why we need an open standard for the multi-agent future. We need a dependable way for agents to interact with one another, to cross boundaries across applications and ecosystems and do it in a way that keeps businesses in control of their product experience.

You might be just as likely to kick off a set of work agents from within an Atlassian Jira ticket connected to a Salesforce customer case as you would want to kick off a set of agents in reverse originating from within Salesforce connected to Atlassian. For agents to work together regardless of where a work request originates and in any number of directions with a consistent user experience, again a standard for doing this is needed.

What else should be represented in this standard? Outside of Salesforce, the multi-agent ecosystem today is an exciting wild west. On the daily we see new innovations and ways of connecting and constructing AI systems and agentive workflows. One recent tie-in between the AI framework LangChain and a tool called Assitant-UI brought this insight:

“UX is crucial for agents. Everyone wants agents with streaming, generative UI, and human-in-the-loop in their application.”

Indeed, we’ve already covered how crucial user experience is to agents. And clearly agents must be able to quickly stream their responses when working with other agents. But what of generative UI and human-in-the-loop in their application?

Let’s start with human-in-the-loop; another area of broad agreement. While Salesforce and others talk a big game about automation, it’s always grounded in the need to be able to bring a human back into the center when necessary. We learned this lesson as well at Clockwise, and have built our scheduling agent experience around a core concept of being able to check back in with the user with a proposed set of scheduling options. When you are doing complex work, it’s amazing to get to full automation, but it starts on the backbone of involving the user and keeping them in the loop. Any standard must be built around an optional ability to check-in and confirm with the user before proceeding, and eventually allow for full automation when confidence is high enough.

And so what of generative UI? Here I’d propose that what is needed isn’t necessarily generative UI but “native UI.” What is important is that the agent is producing a UI that is native and controlled by the service/agent responding to the request. Only the native service will have the context and understanding necessary to render a user interface that ties into the agent request. Whether that UI is rendered using generative AI or some other non-AI mechanism is left to the responding service as an implementation detail. And so here, we think the open standard must allow for the responding service to control and deliver native UI to an agent request.

What comes next? We’re excited to continue to examine what an open multi-agent future might look like. We’ve created a draft of something we’re calling Open Multi-Agent Protocol (OMAP) and we’re excited to keep pushing the conversation forward. It won’t be long before there are entirely new types of jobs out there where people use agents to do work in powerful and streamlined ways. The age of the Agent Orchestrator job description is upon us, and while Salesforce paints a compelling path ahead, we’ll need a standard way for agents to interconnect beyond boundaries.

Source Link

Related Posts

Leave a Comment