Last Week in AI: 19–25 October 2026

19–25 October 2026: changes worth putting into an IT work plan

This week’s briefing looks at a Claude integration upgrade, Google’s explanation of its Gemini customisation hub and a change in the stewardship of open agent projects. The useful question is what each development changes for a system or rollout you already own.

We have reviewed the primary announcements, not tested the products. The practical exercises below are editorial suggestions; the business example is fictional. Dates distinguish the announcement from the rollout where those differ.

21 October: treat the Claude SDK upgrade as an application change

Anthropic released API version 2026-10-21 and new SDK and command-line client versions that select it by default. Its release notes also flag breaking client changes independent of the API version. For an organisation running an internal Claude integration, this is a reason to review a dependency update before promoting it.

Our suggestion: ask the application owner to identify the installed client version and the API version actually sent by the application. Put both in the change record. Have the developer read the linked migration guide rather than assuming that keeping one setting unchanged preserves every behaviour.

Consider a fictional distributor whose internal assistant drafts stock enquiries. Build a small test set from invented requests with known expected results. Include an empty result, an unavailable upstream service and a response that requires a person to decide what happens next. Compare the proposed upgrade with the existing application and investigate differences before rollout.

Agree who can reverse the deployment and what evidence would trigger that decision. A clean installation is only the start of acceptance; the business owner should review the output that colleagues will actually use. Keep any customer-facing sending step outside the trial.

21 October: Google explains where Gemini customisation is managed

Google described a hub combining skills and connected apps in Gemini. Its announcement says the menu’s gradual rollout began on 13 October: this is a new explanation published last week, not a claim that the rollout started then. It also says most connected apps are currently English-language offerings in the US, with selected availability elsewhere.

The UK planning implication is straightforward: confirm the particular connector in the account and region you intend to use before promising it to staff. The hub’s general availability statement does not establish that every listed service is available to your organisation.

Google points administrators to Apps > Additional Google services > Gemini in the Admin console to review integrations. Our suggested exercise is to choose one proposed connection, document its business purpose and identify who will maintain it. Ask a pilot user to demonstrate the access requested with non-sensitive sample material.

Write a short acceptance note describing the intended task and the data the user expects to make available. If the required connection is absent, keep the training plan focused on the existing workflow rather than presenting a US demonstration as evidence of UK access. This follows last week’s skills timetable coverage with a distinct focus on connection management and regional limits.

22 October: four agent projects move under shared stewardship

The Linux Foundation announced Agent Skills, Open Responses, OpenEnv and OpenSandbox as new Agentic AI Foundation projects. The contributions come from Anthropic, OpenAI, Hugging Face and Alibaba respectively. This is a governance development, not the first release of all four tools.

The announcement describes Open Responses as a specification for interfaces spanning model providers, and OpenSandbox as a platform for running AI applications in sandbox environments. Those are relevant directions for teams considering portability and execution controls, but foundation membership does not prove that a particular deployment is interchangeable or secure.

For your next supplier discussion, ask for a demonstration of the exact workflow you might need to move. Provide a small synthetic task and an expected result, then ask the supplier to show which components would change if you selected a different model provider. Record unsupported features alongside successful steps.

If a supplier proposes a sandbox, ask them to demonstrate what the agent can read, write and reach over the network in that configuration. Decide which failure would stop the pilot. Treat a successful test as evidence for that configuration and task, not permission to give every agent broader access.

Leave the week with a decision someone owns

Choose the item that affects an active plan: an application upgrade, a staff integration request or a supplier evaluation. Produce a short record showing the current state, the evidence needed to proceed and the person responsible for the next decision.

If none affects your environment, retain the dated primary links for the next relevant review. There is no benefit in starting an AI project solely because an announcement appeared.