Agents SDK v1.0.0-alpha.1 was published on November 20, 2026 at 20:51 UTC. It is a preview from the feat/simple-codex branch, and its 1.0.0a1 wheels are distributed only through GitHub Releases for Linux x86-64 and ARM64, macOS on Apple silicon and Windows x64, with a published SHA-256 manifest.
Codex becomes a provider, not a separate server
Passing RunConfig(model_provider=CodexProvider()) to the Runner routes model calls through Codex. The release says the wheel embeds unmodified Codex Core rust-v0.166.0 through PyO3, so calls skip codex app-server startup and JSON-RPC. According to the Codex provider guide, the Python Runner keeps the agent loop, function tools, handoffs, approvals, guardrails and output validation, while Codex supplies the model connection and native workspace tools.
Users of the earlier harness preview (v0.24.0-alpha.8) are told to move from Agent.run() or Session.submit() to Runner.run(), finish pending native runs first, and start fresh sessions because the checkpoint formats differ.
Two approval systems with different failure behavior
The guide is unusually direct here. Native Codex approval prompts are separate from SDK tool approvals, and the provider cannot answer them: a native approval request fails the model call. SDK tool approvals still pause through the Runner's human-in-the-loop flow. For non-interactive read-only work, the docs suggest approval_policy="never" with sandbox_mode="read-only", under which commands that need more permission simply fail.
Two other details matter for deployment. Codex Core runs inside the Python process, so the docs note configured hooks can see the application's environment; restricting what commands inherit via shell_environment_policy does not restrict the runtime itself. And Codex authentication does not configure the separate clients used for tracing or session compaction.
What this means for harness design
Analysis: removing the server hop is a latency and packaging win, but it also removes a process boundary. Secrets in the Python environment are now one layer closer to native command hooks. A run can also stop for two different reasons that look similar in logs: an SDK approval pause, which is resumable, and a native approval failure, which ends the call.
Original exercise, not executed: give a Belgavi's Book Store returns agent one SDK tool, issue_refund, guarded by SDK approval, and allow one native shell command that needs write access under a read-only sandbox. Run CASE-6031 and record which stop occurs first, whether the run state is resumable, and what the trace shows. Jenya reviews whether the two outcomes are distinguishable without reading the code.
Sources accessed November 22, 2026. This is a preview release; no wheel was installed for this digest. Views are my own and do not represent any employer, past or present.
Sources and evidence
Agents SDK 1.0.0-alpha.1 embeds Codex Core in-process as an opt-in provider. Native Codex approvals end the model call while SDK approvals pause it, so log and handle those two stops separately.