Parse the answer.
Keep the evidence.

The SDK now separates the text used for structured-output parsing from the complete response.

OpenAI Responses can emit multiple assistant messages containing JSON. Joining them can break parsing even when the final answer is valid. This patch lets OpenAI and Azure Responses models select the last message marked final_answer for non-streaming parsing while retaining generated content, annotations, conversation history and token accounting.

Evidence & limits. The release and linked code change document the fix and added tests. The change covers generateText, generateObject and related middleware; streaming is unchanged. We inspected the implementation evidence, but did not run the tests or establish how often this failure occurs in production.

The UK’s Budget preview links AI adoption with evaluation capacity and domestic supply.

HM Treasury announces another £120m for the AI Security Institute, a £60m National Cyber Shield including a £25m challenge fund, and a £70m vulnerability centre. It also adds £60m to the Sovereign AI Fund and £60m to sovereign R&D procurement. An agentic-cyber regulatory sandbox is being scoped, not presented as an effective regime.

Evidence & limits. This is an official funding announcement, not proof of disbursement, operational readiness or measured security improvement. The page gives an October 24 date but no time; precise inclusion in the Eastern-time window cannot be established.

Agent Orchestrator’s nightly release preserves resolved settings when retrying a failed task start.

The retry fix retains model, reasoning effort and permissions rather than silently taking changed project defaults. Saved empty settings preserve provider defaults; reasoning effort is checked against the current model catalog. A companion documentation correction distinguishes an ordinary ao send, which queues another turn, from ao send --steer, which reaches an active Chat turn.

Evidence & limits. This is an unstable nightly, not a stable release recommendation. The steering change corrects documentation, not runtime behavior. The retry PR reports passing remote checks but unsuccessful full local Windows suites. These are maintainer reports, not independent validation.