Three frameworks. Dozens of GitHub threads. Hours spent debating LangChain vs CrewAI vs AutoGen vs the OpenAI Agents SDK. A practitioner who has shipped 30 AI agents to paying customers over the past six months argues most of that debate is wasted energy.
The claim is blunt: framework choice has almost no bearing on whether your agents survive contact with real users. Pick whichever one your team already knows, and move on.
The Framework Debate Is a Vibes War
LangChain, CrewAI, AutoGen, and OpenAI's Agents SDK all do roughly the same thing: chain model calls, define tools, and manage multi-step workflows. The differences between them are real but marginal at the level that matters for most production systems. Developers argue about abstractions; agents die in production for other reasons.
The things that actually kill production agents are infrastructure problems that don't make for exciting conference talks. Tool reliability is one - an agent that can't handle a 429 rate-limit error gracefully will fail every time traffic spikes. Error recovery is another - agents without clear fallback behavior when a sub-task fails cascade those failures upward into complete workflow breakdown. State management is a third - most frameworks leave persistent state entirely up to you, which is fine until you realize your long-running agent has no reliable memory between steps.
Context window exhaustion is a quieter killer. An agent that works perfectly in a demo with a short task will slowly degrade as the context window (the amount of text the model can hold in working memory at once) fills up over a multi-step workflow. Most framework comparisons happen in controlled conditions where this never surfaces.
Observability is another gap. When an agent misbehaves in production, you need to know which tool call failed, what the model received as input at that step, and whether the failure was deterministic or random. Building that logging infrastructure is unglamorous work that framework tutorials almost never cover, but it's what makes debugging possible.
What the 30-Agent Sample Suggests
Six months with paying customers is a credible stress test. The argument isn't that frameworks are irrelevant - it's that the selection criterion should be "which one does my team already know" rather than "which one has the most theoretically elegant abstractions."
Teams that stall in framework evaluation cycles often do so because agent architecture feels like a decision where picking the right foundation solves hard problems downstream. It doesn't. The hard problems in production agentic systems are prompt brittleness under real-world inputs, cost overruns from runaway loops, and the difficulty of testing something that behaves nondeterministically by design.
The framework is scaffolding. What you build inside it is the product.