Skip to content
trendoviz

OpenAI Agents SDK v0.25.0 removes Runloop provider

Version 0.25.0 of the OpenAI Agents SDK removes the beta Runloop sandbox provider and changes how Pydantic run contexts are checkpointed. Applications using Runloop or sharing checkpoints across workers need to plan data migration and reader upgrades before moving to this version.

48 out of 100Emerging7D: Insufficient history ProvisionalPublished

Overview

OpenAI released openai-agents-python v0.25.0 on 2026-11-17 [1]. The release notes frame two changes as the main ones: removal of the beta Runloop sandbox provider and a change to Pydantic run-context checkpointing [1]. The repository describes the project as "A lightweight, powerful framework for multi-agent workflows," which is a vendor claim rather than an independent assessment [1].

Runloop users are told to preserve workspace data before switching providers, and provider-specific sessions and snapshots require a separate data migration. The notes suggest migrating to another supported provider such as Modal or remaining on 0.24.0 during migration [1]. Applications that share Pydantic run-context checkpoints across workers are advised to upgrade readers first [1].

The checkpointing change preserves default values and in-place edits in Pydantic checkpoints, while process-local resources now require a custom serializer [1].

The release notes also list several runtime and configuration changes: nested agent-tool approvals can be resumed with output guardrails using RunState schema 1.22; Runner handoffs can opt into recovery through failure_error_function, with exceptions still propagating when it is omitted; the demo REPL can approve or reject individual tool calls; MCP v2 Streamable HTTP supports concurrent requests while MCP v1 continues to serialize them; failed and incomplete Responses calls now retain provider-reported token usage; and utilities for third-party session adapters are exposed [1].

Two model defaults change. OpenAIResponsesCompactionSession now defaults to gpt-6-luna, with model="gpt-4.1" retaining the previous choice, and agent model settings follow the selected model more closely. Native Chat Completions tool requests for gpt-6-sol and gpt-6-luna default to non-reasoning mode when effort is not supplied [1]. These are silent behavior changes that can alter application output and usage characteristics after an upgrade [1].

Dependency updates include mcp from 2.0.0 to 2.3.0, openai from 3.29.0 to 3.48.0, and pydantic from 2.14.2 to 2.14.3; other development dependencies were also updated [1].

Why it matters

Two operational consequences stand out. First, Runloop workloads carry switching costs that are separate from an ordinary dependency upgrade: workspace data has to be preserved, and provider-specific sessions and snapshots need their own migration step, with staying on 0.24.0 offered as an interim path [1]. Second, checkpoint readers should be upgraded before writers, which turns a dependency bump into a staged rollout for teams running shared checkpoint storage [1].

The model-default changes are quieter but affect behavior without a code edit. OpenAIResponsesCompactionSession switching its default to gpt-6-luna, and native Chat Completions tool requests for gpt-6-sol and gpt-6-luna defaulting to non-reasoning mode when effort is absent, can change application output and usage characteristics on upgrade; setting model="gpt-4.1" retains the previous compaction choice [1].

Opportunities

The release notes describe an opt-in boundary between handled and surfaced handoff failures: Runner handoffs can recover through failure_error_function, and omitting it preserves exception propagation [1]. For tool servers, applications using MCP v2 Streamable HTTP can handle concurrent requests while MCP v1 continues to serialize them; the notes provide no throughput or latency measurements, so the practical size of that difference is unknown [1].

The notes also describe nested agent-tool approvals that can be resumed with output guardrails using RunState schema 1.22, a demo REPL that can approve or reject individual tool calls, and exposed utilities for third-party session adapters [1]. The last of these is an extension point for custom session backends, and the release notes are the only source for it [1].

Retaining provider-reported token usage on failed and incomplete Responses calls may give teams a more complete usage record for non-standard outcomes; whether that improves billing reconciliation is a hypothesis, since no cost data is provided [1].

Risks

Checkpoint compatibility is the clearest deployment risk. Applications sharing Pydantic run-context checkpoints across workers are told to upgrade readers first, process-local resources now need a custom serializer, and the notes do not establish whether existing checkpoints remain readable by older versions [1].

The Runloop removal affects deployments that selected it as their execution environment, and the notes do not quantify the migration effort involved [1].

Model defaults also shift without an explicit configuration change: OpenAIResponsesCompactionSession moves to gpt-6-luna unless model="gpt-4.1" is set, and native Chat Completions tool requests for gpt-6-sol and gpt-6-luna default to non-reasoning mode when effort is not supplied [1].

The release notes list an opt-in handoff recovery mechanism rather than describing how it changes default error behavior in existing deployments, and the impact of MCP concurrency support on production tool servers is not quantified [1].

Context

The release changes the operational contract around where agents run and how their state is stored. Runloop was a beta sandbox option in the SDK; its removal means provider selection is no longer fully pluggable for existing deployments, because workloads, sessions, and snapshots tied to that provider have to be moved [1]. The checkpointing change touches a separate layer: Pydantic run contexts that are persisted and then read by multiple workers, a pattern common in distributed agent services [1].

The release notes do not establish whether existing checkpoints remain readable by older versions, so mixed-version worker deployments are an area of uncertainty rather than a documented incompatibility [1].

The model-default changes sit in the same release but affect a different group of readers: applications using OpenAIResponsesCompactionSession or issuing native Chat Completions tool requests for gpt-6-sol or gpt-6-luna, where an unspecified effort or model now resolves differently than before [1].

Known, claimed, inferred

Known: v0.25.0 was released on 2026-11-17 and is compared with v0.24.0; the release notes identify the Runloop provider removal and the Pydantic run-context checkpointing change as the two key changes [1].

Inferred: because provider-specific sessions and snapshots require a separate migration and checkpoint readers should be upgraded before writers, agent orchestration state is partly coupled to both sandbox backend and SDK version. Upgrades for these deployments are therefore data-lifecycle work rather than a dependency-only change [1].

Inferred: preserving token usage on failed and incomplete calls suggests the SDK is making usage reporting more complete for non-standard outcomes, which would help teams account for usage that previously went unrecorded; this is a hypothesis, and the notes provide no cost figures [1].

The repository's description of the SDK as lightweight and powerful is a vendor claim, not an independent evaluation [1].

Limitations & uncertainties

Only official release notes were available for this page. No independent test report, performance evaluation, or quality, latency, or cost measurements were provided for v0.25.0 [1]. The notes do not quantify the effort required for the Runloop migration or the likelihood of checkpoint incompatibilities, and they do not state whether existing checkpoints remain readable by older versions beyond the instruction to upgrade readers first [1]. The page does not verify claims about the availability or maturity of the replacement models [1]. One release-note entry is truncated and another contains only the first highlight, so the full change set is not visible [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementOpenAI released openai-agents-python v0.25.0 on 2026-11-17, and the release notes identify the removal of the beta Runloop sandbox provider and a change to Pydantic run-context checkpointing as the key changes.[1]
  • Primary-source or vendor statementThe release notes advise Runloop users to preserve workspace data before switching providers, to migrate provider-specific sessions and snapshots separately, and either move to another supported provider such as Modal or remain on 0.24.0 during migration.[1]
  • Primary-source or vendor statementApplications that share Pydantic run-context checkpoints across workers are advised to upgrade readers first, and process-local resources now require a custom serializer.[1]
  • Primary-source or vendor statementOpenAIResponsesCompactionSession now defaults to gpt-6-luna, with model="gpt-4.1" retaining the previous choice, and native Chat Completions tool requests for gpt-6-sol and gpt-6-luna default to non-reasoning mode when effort is not supplied.[1]
  • Primary-source or vendor statementMCP v2 Streamable HTTP supports concurrent requests while MCP v1 retains serialized requests.[1]
  • Primary-source or vendor statementRunner handoffs can opt into recovery with failure_error_function, and omitting it preserves exception propagation.[1]
  • Primary-source or vendor statementFailed and incomplete Responses calls now retain provider-reported token usage.[1]
  • Primary-source or vendor statementThe repository describes openai-agents-python as "A lightweight, powerful framework for multi-agent workflows."[1]

Sources

  1. 1
    openai/openai-agents-python v0.25.0
    ●github.comretrieved Nov 18, 2026

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.