Salesforce's Agentic Enterprise Architecture for AI Agents
This article delves into Salesforce's Agentic Enterprise Architecture, a layered framework designed for building and deploying AI agents at scale within enterprise environments. It highlights the architectural components and operational lessons learned from over 20,000 deployments, emphasizing the shift in effort from pre-launch development to post-launch continuous improvement and the critical importance of robust trust and security layers for AI agents.
Read original on ByteByteGoThe Agentic Enterprise Architecture Overview
Salesforce's Agentic Enterprise Architecture provides a structured approach for integrating AI agents across various business operations. It comprises four distinct layers, each addressing a specific aspect of agent functionality and interaction, along with a cross-cutting trust layer. Understanding this architecture is crucial for designing reliable and scalable AI-driven systems.
- Engagement Layer: Handles user interaction with agents through existing tools like Slack, chat, or messaging apps.
- Agent Layer: The core where AI reasoning, decision-making, monitoring, and orchestration of agents occur.
- System of Work: Integrates trusted business applications (e.g., CRM, ERP) where actual tasks like resolving support cases or updating sales pipelines are performed.
- Context Layer: Provides agents with the necessary data and metadata to ground their actions in real-world business context, ensuring informed decisions.
- Trust Layer: Spans the entire stack, supporting multiple LLM providers and enforcing critical guardrails for security, data privacy, and ethical AI use.
Paradigm Shift: Post-Launch Dominance in AI Agent Development
AI Agents: 90% of work is AFTER launch
Unlike traditional software development where most effort is pre-launch, AI agents demand significant post-launch investment. The non-deterministic nature of LLMs means continuous monitoring, iteration, and refinement are essential to maintain consistent and reliable behavior, especially in high-stakes enterprise environments. Launch is the starting line, not the finish line.
This fundamental difference necessitates a shift in development strategy. Successful teams prioritize building an agent that can be iterated upon quickly, dedicating the majority of their budget and effort to continuous improvement post-deployment. This involves meticulous review of agent transcripts, identifying decision errors, and updating instructions, tools, and data sources to enhance performance against defined KPIs.
Designing for Trust, Security, and Safety
Building enterprise-grade AI agents requires robust trust and security mechanisms due to the flow of sensitive data through LLMs. Salesforce emphasizes both input and output guardrails to mitigate risks such as data leaks and hallucinated actions.
- Input Guardrails: Protect data before it reaches the LLM. Key strategies include: Secure data retrieval (controlled access to data), zero data retention (LLM providers do not store or use prompts/responses for training), and keeping data inside a trusted boundary (hosting models within your platform's security perimeter for ultra-sensitive workloads).
- Output Guardrails: Validate the LLM's response before it reaches the user to prevent erroneous or harmful actions. This ensures agents operate within predefined deterministic workflows while leveraging probabilistic AI for adaptability.