What should I use to build an agent that can act in Salesforce?

To build an agent that can act in Salesforce, you generally want this stack:

• An LLM agent framework: to reason, plan, and choose actions
  Examples: LangGraph, OpenAI Agents SDK, Microsoft AutoGen, CrewAI

• Salesforce APIs / integration layer: to actually perform actions in Salesforce
  Most common options:

  • Salesforce REST API for CRUD on records
  • Salesforce Bulk API for large data operations
  • Salesforce Tooling / Metadata APIs for config/dev tasks
  • Apex / Flow / Platform Events if you need Salesforce-native automation

• A secure tool-calling layer: expose only the actions the agent is allowed to do
  For example:

  • create_case
  • update_contact
  • search_accounts
  • log_task
  • send_email

  Each tool should validate inputs, enforce permissions, and log actions.

• Auth and permissions:

  • Use OAuth 2.0 connected app / service account
  • Apply least privilege
  • Prefer a dedicated integration user
  • Consider field-level and object-level access rules

• Guardrails / human approval for risky actions:

  • Require confirmation before deleting records, sending emails, changing ownership, etc.
  • Add policy checks and audit logging

Recommended practical setup

If you want the simplest robust approach:

• LangGraph or OpenAI Agents SDK for orchestration
• Salesforce REST API wrapped as tools
• OAuth connected app for authentication
• A policy/approval layer for sensitive actions

If you want the agent inside Salesforce

Use:

• Salesforce Flow + Apex for native workflows
• Or a service running externally that Salesforce calls via API/webhooks

Best choice by use case

• Basic CRUD assistant: LLM + Salesforce REST API
• Complex multi-step workflows: LangGraph + tools + approval steps
• Enterprise governance-heavy use: external orchestrator + Salesforce APIs + audit/policy engine

If you want, I can sketch a reference architecture or give you a minimal Python example for a Salesforce agent.