Solving LLM Challenges: Agentforce’s AI Approach to Reliable Agent Recommendations

In our “Engineering Energizers” Q&A series, we explore into the stories of exceptional engineering leaders driving transformative innovation. Today, we highlight Annie Zhang, Software Engineer Architect on Salesforce’s AI Cloud team. Annie is at the forefront of advancing Agentforce, Salesforce’s AI for autonomous task management. Her team is developing cutting-edge technology designed to deliver personalized, intelligent recommendations that elevate interactions across both customer-facing and employee-facing agents.

Discover how the AI Cloud team addresses the non-deterministic nature of large language models (LLMs). Learn how they scale systems effectively with modular microservices architecture and balance rapid deployment with trust and security to deliver high-performance solutions.

What is your team’s mission?

The team’s mission is to revolutionize user interactions by delivering personalized agent recommendations. Our goal is to guide users effectively through their workflows, whether they engage with Copilot, employee-facing, or customer-facing agents. These recommendations serve as intelligent prompts, suggesting the “next best action” to simplify decision-making and make interactions seamless.

For example, if a user updates their order address, the system might suggest follow-ups such as summarizing the order details, providing an overview of the order status, or checking the estimated delivery date. By dynamically generating these recommendations based on real-time contextual data, the system eliminates guesswork in user navigation.

This process streamlines interactions by leveraging advanced algorithms that bridge user intent with actionable insights. The team ensures every interaction feels intuitive, efficient, and personalized.

Annie describes the engineering culture at Salesforce.

What was the most significant technical challenge the team faced recently?

Interacting with large language models (LLMs) presented significant challenges due to their non-deterministic nature. While these models are powerful, they don’t guarantee identical outputs for the same input, which posed a critical issue: how can we ensure consistent, reliable recommendations for users?

The primary hurdle was tackling hallucination, where models might generate inaccurate or irrelevant responses. To mitigate this, the team conducted exhaustive testing to identify patterns of inconsistency. Fine-tuning the models was pivotal, aligning outputs more closely with user expectations. Additionally, safety layers were implemented to filter harmful or biased content, ensuring ethical AI usage.

For instance, when users reported discrepancies in repeated queries, the team developed iterative testing protocols to stabilize responses. This approach required innovation and persistence to refine the system and deliver dependable AI recommendations.

How are scalability challenges in Agent Recommendations addressed?

In one instance, the team faced a sudden traffic surge on a critical service. By isolating and scaling this service independently, they resolved the issue without disrupting overall system performance. This demonstrated the effectiveness of the modular architecture in maintaining reliability under high demand.

How does the team ensure enhancements in one area of Agent Recommendations do not compromise other areas?

For example, during the rollout of enhanced recommendation algorithms, integration tests revealed conflicts with caching systems. Cross-team collaboration resolved these issues, maintaining performance benchmarks. This structured approach guarantees that platform enhancements are robust and regression-free.

How does the team balance the need for rapid deployment with maintaining high standards of trust and security?

It requires a meticulous approach. When customer support issues arise, these are prioritized above all else. For example, if a bug impacts critical functionality, the team responds swiftly, dedicating all resources to resolution. This focus on customer satisfaction ensures trust isn’t compromised.

In one notable challenge, the team had to deploy a high-stakes feature under tight deadlines. By integrating automated testing pipelines with real-time monitoring, potential risks were mitigated swiftly without compromising speed. This balance is vital for delivering trustworthy, high-performance systems.

Annie shares what keeps her at Salesforce.

What ongoing research and development efforts are improving Agent Recommendations’ capabilities?

Current research and development efforts are focused on enhancing personalization in Agent Recommendations, moving beyond just accuracy to create truly engaging user experiences. Key advancements include integrating contextual signals into the system, such as preferred language settings based on user profiles, interaction history to understand recurring needs, and tonal analysis to match the user’s communication style.

By developing real-time processing pipelines for these contextual signals, we enable adaptive and personalized recommendations. For instance, if a user frequently inquires about delivery times in a specific manner, the system adapts to present recommendations in a familiar format. These innovations bridge the gap between generic functionality and user-specific experiences, enhancing both engagement and efficiency.

What strategies are used to gather user feedback on Agent Recommendations, and how does it influence future development?

One pivotal redesign resulted directly from customer feedback about navigating recommendation options. Collaboration with product managers translated these insights into actionable updates that refined usability. This continuous feedback loop ensures the platform evolves to meet user needs effectively.