The Runloop sandbox provider is removed, so projects using it must switch. New session-adapter utilities and opt-in handoff error recovery help with custom persistence and resilient multi-agentUsing several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.Full definition → flows.
Key takeaways · AI-distilled
openai-agents-python 0.25.0 removes the beta Runloop sandboxAn isolated environment where AI-generated code or agent actions run without being able to touch anything real.Full definition → provider. Runloop session IDs and snapshots cannot be resumed by Modal, so users must preserve workspace data and restore it into a new sandbox, or stay on 0.24.0 while migrating.
RunState checkpoints now write schema 1.22, which a 0.24.0 worker rejects. The release review says to upgrade resume workers before enabling new writers, or route checkpoints by SDK version. Pydantic contexts with process-local defaults can now require a custom serializer.
OpenAIResponsesCompactionSession now defaults to gpt-6-luna instead of gpt-4.1; setting model="gpt-4.1" keeps the old choice. Chat Completions tool requests to gpt-6-sol and gpt-6-luna default to non-reasoning mode when no effort is supplied.
MCPThe Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.Full definition → v2 Streamable HTTP now supports concurrent requests while MCP v1 stays serialized; the review advises enabling tool retries only when replaying a call is safe. Failed and incomplete Responses calls now keep provider-reported tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → usage.
Terms in this piece · Glossary
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.