OpenAI published version 2 of its open-source Agents framework yesterday, with a major change: native support for competing models, notably Anthropic's Claude and Google's Gemini. Developers can now orchestrate multi-agent workflows where each agent uses the model best suited to its task.
v2 also introduces a more robust persistent state management system, based on pluggable stores (Redis, Postgres, or custom solution), and a new observability API that allows tracing each model call and routing decision. The framework remains under MIT license and installs via pip.
This repositioning is strategic: by making its framework model-agnostic, OpenAI implicitly acknowledges that the agents market won't be monopolized by a single provider and seeks to establish itself as the standard orchestration layer, even at the cost of letting competitors benefit from the traffic.