Agentforce
The agentic AI product from Salesforce.
What is it?
- A set of tools
- A suite of agents
- Autonomous: AI Agents or “digital workers” that perform tasks in real time
- Can hand off to a human
- Examples: Autonomous Service Agent, Personal Shopper, SDR
- Assist & Augment: Augments humans by helping them perform tasks more efficiently
- Examples: Service Planner, Reply Recs, Case Wrap-Up, Knowledge Creation, Sales Coach
- Autonomous: AI Agents or “digital workers” that perform tasks in real time
Agent Topics
- Topics define what jobs you need your agent to do
- Agents handle assigned topics like “Order Management”, “FAQ”, “Appointment Scheduling”
- Terminology / configuration options:
- Topic Label: A term that categorizes the jobs to be done within the Topic.
- Classification Description: A plain text sentence that will guide the LLM Topic classifier on when to select this Topic.
- Used to classify the current Topic at each turn of conversation, and most effective when it closely mirrors the types of inquiries expected to “trigger” this Topic.
- Scope: The scope is used to determine what agent can do once the conversation is switched to the current Topic, and helps to constrain agent to not attempt to respond outside the desired job or “scope” you want this Topic to include.
- This does not get sent to the Topic classifier, but is included along with actions in the prompt the agent uses to react and reply to customers.
- Used to guide exactly what an agent can and cannot do for a given Topic (“your job is only to…”), and most effective when it narrowly defines an agent’s job for a given Topic.
- Instructions - Plain text sentences to guide the LLM on how to best use the tools and perform the Topic’s job.
- Best practices for Building Topics
- Keep your topic labels 1 word or short
- When writing the classification description, imagine the Agentforce Service Agent (ASA) can only read the description and must use it to figure out if the current conversation is related to the Topic
- When write the scope, imagine ASA can only read the classification description and scope, and must decide what the agent is and is not allowed to do within this topic
- When writing instructions, imagine ASA can read the classification description, scope, actions descriptions, and other instructions, and must now decide to which actions to use and how to apply them to serve the conversation
- Keep instructions to 1 or 2 lines
- Keep them simple
Additional Implementation Best Practices
- Good sweet spot is 6-8 instructions per topic
- More than this will result in hallucinations
- No more than 15 actions per topic
- No more than 15 topics per Agent
- Really stop to consider the consumption model of pricing
- Play with keywords in your instructions, it will yield slightly different behavior
DevOps / Deployment Notes
- There is indeed an automated process to assist from sandbox-to-production deployments, though it’s not fully fleshed out or documented
- As of Jan 2025
- More tools will be launched in Feb 2025 as part of the Agentforce 2.0
- Even when developing in a Sandbox, any request to the Einstein Requests Pool counts against your production credits
- Talk with your AE about getting a bucket of additional request credits for development at a discount or for free.
Miscellaneous Notes
- Agentforce licenses do come with SOME Data Cloud licenses / availability
- Every action taken by Atlas Reasoning Engine is logged in Data Cloud for audit trail purposes
- 2FA Verification for Agents to use when speaking with customers is now available