AI Agent Studio

Built for governed enterprise AI on Salesforce

Open Source Enterprise Ready Salesforce Native

A comprehensive framework for building AI agents that orchestrate LLM interactions, execute tools, manage memory, and coordinate multi-agent workflows—all while respecting Salesforce security and governor limits.

4 Agent Types
Conversational, Function, Workflow, and Email orchestrators.

Trust Layers
PII masking, prompt safety, and dependency validation.

Multi-Provider
OpenAI, Claude, Gemini with extensible adapters.

Repository Scope

This public repository contains the core framework. The aiAgentStudioAddons package includes additional orchestrators, LLM providers, actions, and UI components that are not part of the open source release.

What It Does

AI Agent Studio is a production-grade orchestration system for AI agents on Salesforce. It handles:

• LLM interactions with multiple providers (OpenAI, Claude, Gemini)
• Tool execution with governed access to Salesforce data
• Memory management across multi-turn conversations
• Multi-agent workflows with state machines and coordination
• Email thread processing with automatic reply generation
• Human-in-the-loop approvals for sensitive operations
• Security controls including PII masking and prompt injection detection
• Observability with execution traces, token tracking, and cost analytics

Why AI Agent Studio

Production-Ready Architecture

Built with Strategy + Factory + Chain of Responsibility patterns. Modular plugin-based design scales from single conversations to high-throughput enterprise deployments.

Security & Governance First

Automatic CRUD/FLS/sharing enforcement, PII masking, prompt safety checks, tool dependency validation, and comprehensive audit trails. Agents run in user context with no privilege escalation.

Extensible at Every Layer

Custom actions for business logic, context providers for dynamic data, LLM provider adapters, custom orchestrators, and configurable memory strategies.

Observable by Design

Execution storyboard UI, decision step logging, tool rationale capture, token and cost tracking. Built-in visibility for monitoring reliability and business impact.

Who It Is For

Admins

Configure AI behavior, capability access, and approvals without writing code.

Managers

Monitor reliability, risk, and business impact with observable agent operations.

Developers

Extend with custom actions, context providers, and LLM adapters.

Agent Types

Conversational Agents

Multi-turn chat assistants with memory and tool execution. Perfect for customer service, sales assistance, and interactive troubleshooting.

Function Agents

Single-task specialists for classification, enrichment, and automation. Ideal for record enrichment, sentiment analysis, and automated decision-making.

Workflow Agents

Orchestrated multi-agent execution with state machines and conditional branching. Native Salesforce reporting on workflow step performance.

Email Agents

Email thread processing with automatic continuity and reply generation. Ideal for automated customer support and email triage.

Security & Trust

The framework implements multiple layers of security and trust controls:

Platform-Native Security: Agents run in user context, CRUD/FLS automatically enforced, sharing rules respected, all actions logged.

PII Masking: Hybrid approach using Salesforce Data Classification and regex patterns. Sensitive data masked before LLM calls, automatically unmasked in responses.

Prompt Safety: Multi-layered jailbreak detection using pattern matching, heuristic analysis, and structural analysis. Configurable response modes (Block/Sanitize/Flag).

Tool Dependencies: Declarative sequencing prevents illogical tool order (e.g., sending email before creating record). LLM-generated dependency graph approved by admins, enforced at runtime.

Human-in-the-Loop: Configurable approval workflows for sensitive operations with Confirmation, Approval, or hybrid modes.

Security Deep Dive

Key Features

Performance

Multi-LLM Optimization: Deferred DML pattern enables multiple LLM calls within a single transaction when tools don’t perform DML or callouts. Reduces latency and async job overhead.

Dead Letter Queue: Resilient multi-record processing with priority ordering, exponential backoff, stale job recovery, and trigger-based chaining. Scales to any volume.

Platform Events: High-throughput async dispatch for production environments with thousands of concurrent users.

Observability

Execution Storyboard: Visual UI showing complete agent execution timeline with steps, decisions, and outcomes.

Tool Reasoning: LLM explains why it selected each tool, captured in execution logs for transparency and debugging.

Token & Cost Tracking: Per-step tracking of prompt/completion tokens and estimated costs. Build dashboards to monitor spend.

Decision Logging: Complete audit trail of user inputs, LLM responses, tool calls, and results for compliance and troubleshooting.

Extensibility

Custom Actions: Extend with business logic by creating Apex classes implementing IAgentAction. Full access to Salesforce data and external integrations.

Context Providers: Supply dynamic data to agents via IAgentContextProvider. Bulk-safe queries with security enforcement.

LLM Providers: Add new LLM providers by extending BaseProviderAdapter. OpenAI included, Claude and Gemini in addons.

Memory Strategies: Implement custom memory management via IMemoryManager. BufferWindow and SummaryBuffer included.

Architecture Details

Quick Start

Time to First Agent

You can have a working first agent in about 30 minutes.

1. Deploy the framework to your org.
2. Configure Named Credentials for your AI provider.
3. Create an LLM Configuration.
4. Create an AI Agent Definition.
5. Add capabilities and configure approvals.
6. Place the chat component on a Lightning page and test.

Start with the guided setup and iterate from observability data.

Launch Setup Guide

Continue

Getting Started

Configuration

Developer Guide

Architecture

Security

Actions Reference