Microsoft moved AutoGen to maintenance mode in Q1 2026 and shipped Microsoft Agent Framework (MAF) 1.0 as its greenfield successor in April, according to Digital Applied and Uvik. The community continues the original AutoGen lineage as AG2, an independent open-source fork. Despite 55,000 GitHub stars and an active community, AutoGen will receive bug fixes and security patches only. No new features.
That shift caps a year of rapid consolidation in the AI agent framework market. Three frameworks now dominate production discussions, each with a distinct mental model, and a fourth category has emerged that questions whether frameworks are the right abstraction at all.
LangGraph: The Production Default
LangGraph models agent workflows as directed graphs with typed state schemas, built-in checkpointing, and durable execution. It reached 1.0 general availability in October 2025, and by mid-2026 it has accumulated the strongest production track record of any open-source agent framework.
The numbers are concrete. Klarna’s AI Assistant, powered by LangGraph and LangSmith, serves 85 million active users and reportedly reduced customer resolution time by 80%, according to LangChain’s case study documentation cited by Digital Applied. LinkedIn runs a hierarchical recruiter agent and a SQL Bot on LangGraph. Uber’s Developer Platform uses it for automated unit-test generation, reportedly saving approximately 21,000 developer hours.
The trade-off is learning curve. Graph-based thinking is not intuitive for most engineers. Defining nodes, edges, conditional branches, and state schemas requires a mental model shift that, as DEV Community contributor Cristian Iridon wrote, “takes weeks to internalize.” Teams also bear the infrastructure burden: LangGraph provides orchestration primitives, not a managed runtime. Compute provisioning, authentication, logging, and deployment remain the team’s responsibility unless they pay for LangGraph Cloud.
CrewAI: Fast to Prototype, Harder to Scale
CrewAI takes the opposite approach. Role-based crews with goals and backstories map directly to how non-technical stakeholders think about team delegation. A working multi-agent system can be scaffolded in 15 lines of Python. According to Iridon’s analysis, CrewAI has reached 60% Fortune 500 adoption, backed by Insight Partners, with 44,000+ GitHub stars.
The ceiling appears when workflows need to loop, retry dynamically, or branch based on intermediate results. CrewAI’s sequential and hierarchical process models work for linear pipelines but, as Uvik’s production comparison notes, many teams eventually outgrow the simpler role-based orchestration. There is no built-in checkpointing: if a four-agent crew fails on the third task, the entire crew restarts.
For teams where speed-to-demo matters more than production resilience, CrewAI remains the fastest path from idea to working prototype. For teams operating in regulated environments where auditability and deterministic control are requirements, LangGraph is the safer bet.
The AutoGen Question
AutoGen’s move to maintenance mode creates a migration problem for teams already running it in production. Uvik reports that Microsoft unified AutoGen and Semantic Kernel into MAF, which reached v1.0 general availability in April 2026. The AG2 community fork keeps the original codebase alive, but teams evaluating frameworks for new builds are being steered elsewhere.
Digital Applied flagged a practical consequence: teams that evaluate frameworks from 2024 blog posts are still being sent to deprecated paths. Maintenance-mode frameworks receive bug fixes only, and archived repos receive nothing. The mis-routing has “real consequences” for teams that discover the transition after committing engineering months.
Frameworks vs. Platforms
The more consequential shift may be the emergence of managed multi-agent platforms as an alternative to framework-based development entirely. Iridon’s analysis introduced Progenix and Nexus as examples of platforms that bundle orchestration, observability, governance, and multi-tenancy out of the box, rather than requiring teams to assemble those capabilities from separate tools.
The Uvik production comparison reinforces this framing: framework choice in 2026 can change agent performance dramatically even when the underlying model stays the same, which means orchestration is not a secondary implementation detail. Teams that treat framework selection as a footnote to model selection are making the wrong bet.
The pattern echoes a familiar infrastructure cycle. Container orchestration went from manual Docker Compose files to Kubernetes to managed Kubernetes (EKS, GKE) in roughly five years. AI agent orchestration appears to be compressing a similar transition into half the time. The question for engineering teams is not which framework to adopt. It is whether the framework era is already ending.