Framework Watch: OpenAI's Agents SDK Alpha Makes Strict Writes the Default, and ADK Java Stops Discarding MCP Results

Most of this week's framework news was covered in earlier digests. These items are new, and the first is explicitly an alpha, so it should be read as a signal of direction and not as something to deploy.

OpenAI Agents SDK 0.24.0-alpha.7

Published on November 8, 2026, this pre-release changes several defaults that would catch an unwary upgrade.

Strict writes. New Environments default to write_policy="strict", which requires a compatible Linux or macOS executor. The working directory remains writable, and writable_roots adds others. Windows executors need write_policy="legacy", and restoring a legacy checkpoint requires choosing legacy explicitly, because reopening does not upgrade the policy.
Reasoning effort. The bundled catalog changes the default from low to medium for gpt-6-astra, gpt-6.1-sol and gpt-5.6-sol. An upgrade could therefore raise cost without any change in application code.
MCP approvals renamed. "auto" becomes "native" and "approve" becomes "preapproved". The old values are rejected.
Application-owned history. New SessionStore, SQLiteSessionStore and CustomStorage persist root history and automatic Python tool results. The notes limit this first profile, which excludes native tools, Environments, children and generated-code execution. Stores can report uncertain root calls for reconciliation, and "neither automatically replays those effects".

That last point deserves attention. A tool call whose outcome is unknown should not be repeated automatically, and it is useful to see a framework say so in its release notes.

ADK Java and smaller changes

MCP result mapping. A November 8 commit on ADK Java's main branch adds an opt-in CallToolResultMapper. By default, MCP tools return only the text content under text_output. A mapper set on the toolset builder can return other parts of the result, such as structuredContent or images. The interface notes warn that the response is sent to the model, so it must be non-null and JSON-serialisable, and that a mapper exception fails the call without retry.
Nullable fields. A second November 8 commit keeps the null of a required nullable property in set_model_response. Both changes are unreleased, since v1.14.0 dates from November 2.
Skills licensing. A November 6 commit moved the docx, pdf, pptx and xlsx skills in anthropics/skills to Apache 2.0, matching most of the repository. Teams that had avoided those files on licensing grounds should read the new terms.
Gemini CLI. A November 6 change allows overriding the 60-second API headers timeout across all authentication modes, through model.requestTimeoutMs or GEMINI_API_HEADERS_TIMEOUT_MS.

What remains uncertain. We read the release notes, commit messages and one interface file, and did not run these changes. An alpha may change again before release, and the ADK commits may ship differently.

Analysis: Pin model settings explicitly, because a framework's defaults can change the bill. If Jenya's returns assistant at Belgavi's Book Store relied on implicit reasoning effort, a catalog update could change its cost and latency with no code change. Before opting into a richer MCP mapper, decide which fields the model should see, since everything a mapper returns becomes model input and enters the same trust boundary as any other tool output.

Sources and evidence

OpenAI Agents SDK 0.24.0-alpha.7 defaults new environments to strict writes, unavailable on Windows executors, raises default reasoning effort for three models, and adds application-owned history that will not replay uncertain tool calls. ADK Java's main branch adds an opt-in mapper for full MCP results. Pin model settings explicitly and decide what tool output reaches the model.