Unified Planner: Agentforce AI Execution Engine

Unified Planner: Agentforce's AI Execution and Reasoning Engine

Unified Planner is the new AI execution and reasoning engine powering Agentforce, designed to unify previously separate AI agent runtimes across voice, text, and chat experiences. This architectural shift significantly reduces response latency and provides a common runtime for developers to build and manage agents.

Mission and Mandate

The core mission is to provide a singular AI runtime capable of supporting all Agentforce interactions, irrespective of their origin (voice, chat, text, or future modalities). Unified Planner acts as the central execution and reasoning layer, enabling developers to create agents once and rely on a common runtime for execution, observability, memory management, tool orchestration, and model integration.

This mandate presents several challenges:

• Low Latency: Voice interactions demand minimal response times.
• Complex Workflows: Enterprise processes require deep reasoning and orchestration capabilities.
• Platform Independence: Integration with MuleSoft and other platforms necessitates execution capabilities without Salesforce-specific dependencies.
• Future Modalities: Emerging interaction patterns, such as video, introduce new complexities.
• Extensibility and Consistency: Customers expect platform extensibility, while platform teams require operational consistency.

Simple solutions often optimize for one constraint at the expense of others. Historically, systems optimized for speed might sacrifice flexibility, while those focused on reasoning could introduce latency. Separate runtimes address immediate problems but lead to long-term fragmentation.

Addressing Previous System Limitations

Prior to Unified Planner, distinct execution systems had evolved independently. Agentforce's Agent Graph focused on reasoning and orchestration, while Voice Planner was optimized for low-latency interactions. This divergence led to duplicated capabilities, architectural assumptions, and operational models, making maintenance costly and slowing innovation. Developers experienced inconsistent capabilities based on the runtime used, leading to a less predictable experience.

To overcome this, the team separated platform-managed concerns (e.g., prompt injection detection, interruption handling, execution infrastructure, AI runtime services) from customer-defined business workflows. This separation facilitated the unification of previously isolated runtimes while preserving specialized functionalities, establishing a common execution foundation for a more consistent development and operational experience across Agentforce.

Latency Reduction: From 20 Seconds to 2.3 Seconds

A primary architectural change was the aggressive adoption of parallel execution. Operations that were previously sequential were redesigned to run concurrently where dependencies allowed. Tasks such as prompt injection detection, citation generation, grounding validation, knowledge retrieval, and context gathering, which individually added latency, were refactored.

The AI execution engine was redesigned to allow platform services to execute in parallel. This capability was extended to customer workflows through configurable parallel tool execution. Multiple tool calls that once ran sequentially can now execute simultaneously when dependencies permit.

Model selection was also optimized. Developers can now choose models aligned with specific latency and reasoning requirements, utilizing lightweight internal classifiers for specialized tasks. These architectural enhancements reduced average response times from approximately 20 seconds to around 2.3 seconds, making real-time AI interactions significantly more viable.

Extensibility Challenges for a Unified Engine

Designing a single execution engine for Agentforce, MuleSoft, and future interaction modes presented extensibility as a core challenge. The design had to accommodate requirements that were not yet defined, such as stringent latency constraints for voice, sophisticated orchestration for Agent Graph, and independent operation for MuleSoft.

Instead of a monolithic system optimized for current needs, the focus was on creating reusable execution primitives that support diverse reasoning patterns without dictating a single implementation strategy. The core AI execution engine remains shared, with individual products extending it through integrations, policies, and platform-specific services. This architecture allows Agentforce and MuleSoft to share the same execution foundation while maintaining flexibility for different deployment models, customer requirements, and future innovations.

Safely Migrating Production Agents

The migration of production agents onto Unified Planner required careful execution to avoid disrupting customer experiences. Agents had evolved under varying assumptions, prompting strategies, and model behaviors. Due to the probabilistic nature of large language models, even minor runtime changes could lead to observable output differences that customers would notice.

Unified Planner also introduced new capabilities that were not uniformly supported across all clients. Some interfaces could immediately leverage new functionality, while others required additional development. Voice interactions presented a unique challenge, as historical systems treated the end of a call as a session end, whereas Unified Planner introduced session portability across interaction modes.

A successful rollout involved extensive testing, customer validation, selective feature activation, and cross-team collaboration. By approaching the migration as an engineering problem, the team transitioned customers with minimized disruption.

Emerging Challenges: Multimodal Environments

As Agentforce expands into video and other future interaction modes, a significant emerging challenge is extending reasoning systems into fully multimodal environments that support open-ended reasoning. Video introduces complexity through simultaneous, multi-stream information arrival, including visual content, spoken language, contextual signals, and user interactions.

Multimodal experiences necessitate that the AI runtime determines how these diverse signals are represented, interpreted, and integrated into reasoning workflows. Two approaches are being explored:

1. Structured Input Transformation: Continuous streams are transformed into structured inputs before reasoning commences.
2. Multimodal Foundation Models: These models handle signal transformations internally.

Both approaches involve tradeoffs in performance, accuracy, operational complexity, and extensibility. Determining the optimal balance is an ongoing area of research and development.

Key Takeaways

• Unified Planner is Agentforce's new AI execution and reasoning engine, consolidating disparate runtimes.
• It significantly reduced response latency, from ~20 seconds to ~2.3 seconds, by enabling parallel execution of platform services and customer workflows.
• The architecture emphasizes reusable execution primitives, supporting extensibility across Agentforce, MuleSoft, and future modalities.
• Safe migration of production agents involved rigorous testing, customer validation, and cross-team collaboration.
• Future challenges include extending reasoning capabilities into multimodal environments, particularly for video interactions.