TDX 2025 - Day 1 (Part 2)
Additional notes from the first day of Trailblazer DX, the yearly Salesforce developer conference. See the first part of my day 1 notes here.
Atlas Reasoning Engine: Under the Hood
- Systems of thinking
- System 1 Thinking <- Traditional AI is good at this type of thinking
- Instinctive thinking
- Pattern based
- Static
- System 2 Thinking <- Reasoning models tackle this level
- Analytical
- Deliberate
- Effortful
- System 1 Thinking <- Traditional AI is good at this type of thinking
- Atlas Reasoning Engine: The brain behind Agentforce
Atlas Reasoning Loop
- User query is received
- Data Cloud Retrieval occurs
- Initial retrieval first (SFR-Embedding)
- Small tangent on vectors / embeddings
- What does a “good embeddings” mean?
- Numerical representations of text capture the real intent
- SFR-Embedding - Model for embeddings that SFDC has developed through their own research division
- Available via HuggingFace
- Small but very powerful
- Then reranking occurs to get the better (SFR-Reranker)
- Reviews the chunks of text and determines or ranks which ones are best or more relevant
- SFR-Reranker - Near realtime reranking
- Initial retrieval first (SFR-Embedding)
- Planner
- Takes query, breaks it down into manageable steps
- Tool Selection
- Evaluation
- SFR-Judge - Salesforce’s in-house evaluation model
- Feedback
Roadmap Notes
- Note that not all of the models that SFDC has created are in the Atlas Reasoning engine, they’re coming
Agentforce Roadmap
Notes on the things related to Agentforce that are being released, and when they are available.
- Agentforce Consultations
- Ideating with SFDC technical experts
- AI-generated topics, instructions, and actions for your use case
- AI-Assisted Agent Generation
- AI-Generated Actions: June 2025
- Topic Overlap Detection: June 2025
- Agentforce Employee Agent - Roadmapped for when??
- Create task-specific, specialized agents for any department or internal use case
- Topics like employee onboarding, benefits knowledge, general CRM, general FAQ
- Can also answer questions from structured or unstructured data
- AgentExchange
- Available now
- Multi-modal support
- GA April 2025
- Rich Content response support
- GA April 2025
- Voice
- Pilot June 2025
- Agentforce API
- GA Today
- Graph-based Multi-agent reasoning
- On the roadmap, no specific date given
- Supervision Tooling
- Dashboards & Reports: Already GA
- Adoption analysis
- Do I have high-level issues?
- Tracing and Monitoring: “Coming soon”
- Behavior understanding
- What happened in the session? Why is latency high? How did the planner react? Are the right actions picked up?
- Interaction Explorer: “Coming soon”
- Interaction analysis
- What are unresolved conversations? Do I have a knowledge gap? Are customers asking things I didn’t anticipate?
- Agentforce Interaction Explorer
- Complete observability explorer
- Pilot - April 2025
- Dashboards & Reports: Already GA
Testing Center
Build Connected Agents with Agentforce and Mulesoft
- Primary challenges with agents as they exist today
- Agents don’t know how to leverage APIs
- Agents need controlled access to these APIs
- Difficulty accessing APIs when building Agents
- Unsure how to effectively manage Agent growth
- Mulesoft can be used by Agents - Topic Center
- GA Now
- Primary framework for enabling Agents with outside system data via Mulesoft
- Create a relevant Topic
- Create custom Actions that are build on Mulesoft to get the data
- Make sure to provide appropriate instructions and guardrails for when these API operations should be accessible
- Can use Anypoint CodeBuilder to create / design Agent topics
- Make sure to select the “Make Available for Agent Topic” button
Mulesoft for Salesforce: API Catalog
- Unified view of all APIs (Mule, Heroku, Salesforce)
- Requires that you already have linked Mulesoft Anypoint and Salesforce
- From here you can select APIs and activate actions
- Both invocable actions as well as Agent Actions
- Note: Permission sets are created automatically here to give permission to access these APIs
Final Thoughts on Connecting
- In Agent Builder you’ll then find the Topic that you can add directly to an Agent
- Monitoring / Managing
- Use Anypoint Monitoring for this
Securing Agentforce via Sandboxes
- Salesforce Application Lifecycle Management (ALM)
- Plan
- Create
- Test
- Release
- Observe
- repeat…
Recent Sandbox Enhancements related to AI
- Selective Sandbox Access
- Powered by Pubic Groups
- Tightens up access to sandboxes upon creation
- Eliminate need for additional “.invalid” emails
- Required for Dev and Dev Pro sandboxes, but optional for Partials and Full
- Data Mask & Seed
- Product that will remain compliant by deleting or transforming data as it’s pushed into a sandbox
- Comes with a wide variety of masking options
- You can also consider purchasing Own - recent SFDC acquisition
- Recently improved to enhance speed and to be built directly on the core platform
- Data Cloud in Sandboxes - GA Today
- Agentforce Testing Center - GA Today
- Synthetically generated tests for your AI workloads
- Scale Test: Add on to Full Copy Sandboxes
- Temporarily scale up and remove throttle limits to simulate production environments
Demo - Build new Topic / Actions for Agent
- Set up your data masks where needed
- Generate new Dev Sandbox
- Ensure that Data Cloud is provisioned into the Sandbox and validate the connections to external data sources
- Seed the sandbox with data if you need it
- Build new topics, actions (along with flows, Apex, etc…), and instructions
- Navigate to Testing Center (still in sandbox) and test the agent workload
- Release via DevOps Center (or probably still Change Sets too?)
Notes from Day 2
Check out my notes from TDX ‘25 day 2 here.