CrewAI

CrewAI sits in Tony Seale's list of orchestration frameworks alongside LangGraph, AutoGen, Semantic Kernel, the OpenAI Agents SDK, AWS Bedrock and Google ADK, and Seale's point is that the whole category is collapsing. He quotes Anthropic: every component of a harness encodes an assumption about what the model cannot do alone, and those assumptions age fast. Scaffolding built for a limited model handicaps a capable one, so it should shrink over time rather than accumulate.

Cobus Greyling describes the same displacement as a layer collapse with a specific mechanism: natural-language prompts in a CLAUDE.md file took over the orchestration layer that frameworks like LangGraph or CrewAI held, one step after the CLI took the integration layer from MCP.

Antoine Habert reaches the criticism from production rather than from theory. Across two years of agent deployment in a European fintech (2023 to 2025), his AI4Ops system reached 100% automation of level-1 support with more than 90% cost reduction under banking compliance constraints. What that required was decision observability, validated action execution, clear human supervision and auditable feedback. Habert's verdict on the 2024 framework boom, CrewAI included, is that the multiplication of orchestrators left reasoning transparency, structured organizational memory, real cognitive cooperation between agents and scalable supervision underdeveloped. Orchestration was on offer; architectural foundations were not.

For Seale, frameworks are a transitional phase and models are rented. The ontology is what an organization actually owns.