As the “brain” behind Agentforce, Atlas generates a plan based on what the role is trying to do, then evaluates and refines its plan (if required, it can loop to pull in additional data). When taking an action, it will look at what business process (e.g. flows) should be used. It then engages with the customer or employee based on their channel preference.
The Atlas Reasoning Engine is worth looking into further as it’s an agentic system (versus an assistive system), which means it has more agency in actions that it performs on behalf of the user. In this guide, we will take a closer look at what makes Atlas’s new ways of reasoning and planning noteworthy, enabling agents to deliver accurate and relevant results.
What Does Atlas Do?
Atlas is the reasoning and learning engine behind Agentforce, otherwise known as the “chain of thought”. It’s a sophisticated engine that goes in a “flywheel” of retrieving structured and unstructured data (CRM data, data brought in via a zero-copy partner), and then taking action.
Atlas keeps “looping” until it’s confident it can achieve the goal that’s been set by your organization, in line with the user’s ask, whether that user is someone internal in your workforce, or a customer/prospect. Then, it performs “reinforcement learning” based on human feedback – in other words, with every interaction, Atlas will get smarter for your own organization.
From a security standpoint, Altas respects the sharing model as defined at the application layer (e.g. which of these sales reps have access to which customer records?). As explained by one of the leaders close to this processing engine, Atlas is Salesforce’s upgraded copilot engine.
The diagram below is helpful in understanding the core components of the process – the ultimate aim being to generate precise and factual results.
When it comes to Atlas, we can break it down into three capabilities:
1. Atlas Reasoning Engine
As mentioned, Atlas generates a plan based on what the role is trying to do, then evaluates and refines its plan. When taking an action, it will look at what business process (e.g. flows) should be used.
This ability to self-reflect is key, enabling it to be deliberate in its decision-making. Other reasoning engines are more static in nature, and therefore less reactionary to changing customer questions, for example, in the context of a conversation.
2. Advanced Retrieval Mechanics
Also referred to as the grounding mechanisms, this works in conjunction with the reasoning layer. This can be compared to “finding the needle in the haystack” among masses of enterprise data, to then deliver better on the prompts that were given. The overarching aim is to reduce the risk of hallucination – over even, avoid it completely.
3. Guardrails
Another capability that works in conjunction with the reasoning layer are guardrails – or, what agents can’t do. Being able to set guardrails, as in, having policies in place, means that agents are clear on when they are veering out of bounds. The agent would, for example, hand off to a human agent, or ask clarifying questions before going and designing a more dynamic plan.
Summary
Atlas is poised to revolutionize how businesses handle complex processes by offering advanced reasoning, retrieval, and decision-making capabilities. Its potential to transform customer and employee interactions is huge, and we’re likely to hear more exciting updates on its development at Dreamforce.
What are your thoughts on Atlas’s potential? Leave your thoughts in the comments below!
Comments: