We evaluated each platform across six dimensions that matter when agents leave the demo and touch real customers or internal data: deployment flexibility (cloud, self-hosted, VPC, on-prem), LLM support (single vendor vs multi-model), voice capability (real-time phone/voice agents), multi-agent orchestration (role-based, graph-based, event-driven), observability (tracing, cost attribution, audit logs), and governance (RBAC, compliance, data residency). We tested every platform with a live agent that performed a 5-step cross-system workflow (read CRM → qualify lead → send email → log outcome → escalate exception). Pricing reflects the lowest entry point that supports production agents, not free tiers.
Copilot Studio is Microsoft's low-code platform for building, customizing, and deploying AI agents and copilots inside Teams, SharePoint, Dynamics 365, and the broader Power Platform. It targets enterprises already standardized on Microsoft who want conversational agents grounded in internal data without writing code.
Agentforce is Salesforce's native AI agent platform, tightly integrated with Service Cloud, Sales Cloud, and Data Cloud. It deploys autonomous agents for service resolution, lead qualification, and GTM workflows — all grounded in Salesforce CRM data. Agentforce Agent is the conversational runtime; Agentforce Service is the autonomous resolution layer.
Dust is a Paris-built agent platform designed for teams deploying fleets of AI agents across departments. It stands out for multi-model support (GPT-5, Claude, Gemini, open-weight), multiplayer editing, and a knowledge-source layer that connects internal docs, Notion, Slack, and GitHub without data movement. Dust is the bridge between developer flexibility and business-user accessibility.
Vybe is the most ambitious new entrant — when its agents need a tool that doesn't exist (a pipeline dashboard, an invoice tracker), they build a full web app with database, UI, and business logic, then operate it on a schedule. Vybe agents have persistent org memory, per-agent isolation, and full read/write access across integrated tools. It's designed for GTM, operations, and internal tooling.
LangGraph is the stateful orchestration layer from LangChain, and LangSmith is its observability and governance companion. LangGraph's graph-based workflow model (nodes + edges + checkpointing) is the strongest fit for agents that need human-in-the-loop, long-running task execution, and persistent memory. LangSmith adds tracing, evaluation, cost attribution, and compliance dashboards.
CrewAI popularized the role-based multi-agent pattern (agents have roles, goals, and tools). It's the fastest path from prototype to production for teams that want multiple specialized agents collaborating on complex tasks. CrewAI has evolved from a pure framework into a platform with CrewAI Enterprise (AMP) offering governance, observability, and managed deployment.
n8n is the fair-code workflow automation platform that added AI agent nodes in 2026. It combines the visual Scenario Builder paradigm (similar to Make) with self-hosting, code nodes in JS/Python/TypeScript, and MCP support. For teams that need agentic workflows inside a self-hosted or air-gapped environment, n8n is the strongest option.
Rasa is the original open-source conversational AI framework, now evolved into an enterprise platform with Rasa Voice, multi-LLM support, and on-prem/VPC deployment. It's the strongest choice for regulated industries (finance, healthcare, government) that need voice-enabled agents with full data residency and no model vendor lock-in.
| Platform | Deployment | LLM Support | Voice | Multi-Agent | Observability | Governance | Best For |
|---|---|---|---|---|---|---|---|
| Copilot Studio | Cloud (M365) | Azure OpenAI + connectors | Limited | Medium | Medium | Enterprise (Entra) | Microsoft shops |
| Agentforce | Salesforce Cloud | Multi-LLM via Models API | Beta | Medium-high | Medium | Salesforce-native | CRM-centric teams |
| Dust | Cloud / Self-host | Multi-model | Limited | High | High | High | Multi-model teams |
| Vybe | Cloud | Multi-model | In dev | Full | High | High (SSO, RBAC) | Agent workforce |
| LangGraph | Cloud + Self-host | Multi-LLM | Limited | Best-in-class | Best-in-class | Medium-high | Engineering teams |
| CrewAI | Cloud / Self-host | Multi-LLM | Limited | High | Via AMP | Medium | Multi-agent prototyping |
| n8n | Self-host / Cloud | Multi-model via nodes | Via nodes | Medium | Medium | High (self-hosted) | DevOps / self-hosted |
| Rasa | On-prem / VPC / Cloud | LLM-agnostic | Best-in-class | Medium | Medium | Best-in-class | Regulated / voice |
| Platform | Entry Tier | Production Tier | Pricing Model |
|---|---|---|---|
| Copilot Studio | $15/user/mo | Power Automate + Copilot Credits | Per-user + credits |
| Agentforce | Freemium | Flex Credits (~$2/conversation) | Outcome-based |
| Dust | $29/user/mo | $99/user/mo | Per-seat |
| Vybe | ~$500/mo | Custom | Flat-rate team |
| LangGraph | Open-source | LangSmith from $39/seat/mo | OSS + SaaS add-ons |
| CrewAI | Open-source | Enterprise custom | OSS + custom |
| n8n | Free self-hosted | €20–€80/mo | Fair-code + cloud |
| Rasa | Free Developer Edition | Custom enterprise | Open-source + custom |
AI agents have crossed from "cool demo" to "operational infrastructure." The teams that win in 2026 are the ones that choose a platform matching their governance, integration, and deployment reality — not just their model preference.
Frameworks (LangGraph, CrewAI) are developer tools for building agents — they provide the code-level abstractions but leave hosting, observability, and governance to you. Platforms (Copilot Studio, Dust, Agentforce) are production-ready environments that include hosting, multi-model support, observability dashboards, compliance tooling, and often visual builders for non-technical users. In 2026, the line is blurring — most frameworks now offer managed tiers, and most platforms now expose developer APIs.
Not yet, but you should choose a platform with a clear voice roadmap. OpenAI Realtime-2 and Anthropic Claude Voice Mode made voice the fastest-rising agent interface. By Q4 2026, 40%+ of new enterprise agent RFPs will require voice. Rasa leads open-source voice; Retell and Vapi lead managed voice. Copilot Studio and Agentforce are catching up.
Yes, but you need the enterprise tier. LangGraph, CrewAI, and Rasa all offer open-source cores, but regulated industries need managed audit logs, RBAC, data residency, and SLAs — only available in their commercial tiers. Rasa is the strongest for on-prem/air-gapped requirements. LangGraph + LangSmith is the strongest for SOC 2 / ISO 27001 compliance.
n8n self-hosted (free) + open-source LLM (free) + basic hosting (~$5–20/mo) is the cheapest path, but requires engineering maintenance. For managed platforms, CrewAI's open-source core is free; Dust starts at $29/user/mo; Copilot Studio starts at $15/user/mo with existing M365 licenses.
Choose platforms with multi-model support (Dust, LangGraph, n8n) and MCP compatibility. Avoid platforms that require proprietary agent definitions (Copilot Studio, Agentforce) unless you're committed to that ecosystem. Design your agents with model-agnostic prompts and tool definitions from day one — MCP is becoming the universal connector standard.