LangGraph

Antoine Habert, writing in October 2025 after two years of deploying agents in a European fintech, lists LangGraph alongside CrewAI and AutoGen as evidence that 2024 produced orchestration without foundations: no reasoning transparency, no structured organizational memory, no graded supervision. That framing recurs. Tony Seale puts LangGraph at the head of the same list in April 2026 and calls the whole category transitional, arguing that scaffolding built for a limited model handicaps a smarter one and should shrink over time instead of accumulating; the practitioners he watches have moved to a frontier model inside a harness («Claude Code», Codex). Cobus Greyling is blunter, describing an orchestration layer that collapses as natural-language prompts in a «CLAUDE.md» file take over from frameworks like LangGraph.

What survives the critique is narrower and more concrete. LangGraph is the open source demonstration that agent state can be durable, inspectable and portable: checkpointing, human interruptions, resumption after a crash, disposable workers above it. Janakiram MSV credits precisely that in July 2026, then names the catch. LangGraph's control plane sits in «LangSmith», a commercial product. His larger point is that Bedrock AgentCore, Microsoft Foundry and the Gemini Enterprise Agent Platform have converged on one architecture with no portability contract, and that whoever owns the agent control plane gets to define what an agent is. LangGraph serves as both counterexample and illustration: open where the graph executes, closed where the lifecycle is governed.