In this article
Context
We are building agentic features — multi-step AI workflows that plan, use tools, and iterate. We need to decide whether to use LangGraph (stateful graph-based orchestration), build directly on the model APIs with lightweight wrappers, or adopt another framework like AutoGen.
Options Evaluated
LangGraph
Pros
- +Purpose-built for stateful, cyclic agent workflows
- +Built-in persistence, checkpointing, and human-in-the-loop support
- +LangSmith integration provides turn-key observability for agents
- +Active ecosystem with many pre-built patterns
Cons
- −Adds LangChain dependency with its associated abstraction overhead
- −Graph mental model has a learning curve
- −Breaking changes between LangChain versions historically painful
- −Python-first; TypeScript SDK less mature
AutoGen (Microsoft)
Pros
- +Strong multi-agent conversation patterns
- +Native .NET support aligns with our backend stack
- +Microsoft backing provides enterprise-grade support
- +AgentChat abstraction is clean and intuitive
Cons
- −Less mature than LangGraph for production deployments
- −Smaller community and fewer third-party integrations
- −Actor-based model (v0.4) represents a complete paradigm shift from earlier versions
Direct SDK calls with Vercel AI SDK
Pros
- +Zero framework overhead — full control
- +Vercel AI SDK handles streaming, tool calling, and multi-step logic well
- +No dependency on framework versioning
- +TypeScript-first, aligns with our Next.js stack
Cons
- −Must implement persistence, checkpointing, and state management manually
- −No built-in human-in-the-loop support
- −More boilerplate for complex multi-agent patterns
Decision
We chose direct SDK calls with the Vercel AI SDK for our TypeScript/Next.js agentic features. The Vercel AI SDK's generateText/streamText with maxSteps handles multi-step tool use natively, covering 90% of our use cases without framework overhead. For complex stateful workflows (if needed), we will evaluate LangGraph Python microservices. AutoGen is reserved for any dedicated .NET agent services.
Consequences
- •Use Vercel AI SDK generateText with maxSteps for multi-step tool calling patterns
- •Build lightweight state management utilities for complex agent flows
- •Implement manual checkpointing using our existing Neon database if long-running agents are needed
- •Monitor agent costs per run via token tracking middleware — no built-in budget in Vercel AI SDK