Last Week in AI: 12–18 October 2026

12–18 October 2026: the operating details behind AI announcements

This week brings new Copilot Studio capabilities entering a limited rollout, a changed timetable for Google skills and a closer look at independent access to Anthropic’s safety work. Each gives an IT team something more useful than a model leaderboard: a specific question to put into a rollout or supplier review.

We have reviewed the primary announcements; we have not tested these capabilities. The exercises below are our editorial suggestions, and the business examples are fictional.

15 October: Copilot Studio moves towards organisation-owned agents

Microsoft announced agent accounts, business process memory and triggers in the GitHub Copilot harness within Copilot Studio. It says the features are rolling out through its Frontier programme. Treat that as a limited rollout to investigate, rather than assuming every tenant has production-ready access.

The announcement describes agents using their own approved credentials and scoped permissions, retaining process context between cycles, and starting from an event or schedule. Additional controls and automated recommendations for refining agents are described as future work. Do not build a deployment plan around those promised additions as though they are already present.

For a fictional facilities team, try preparing a draft maintenance summary from an approved sample folder. Keep the trial unable to send messages or change service records. Ask the administrator to show which identity performed each action and how that identity can be disabled when the process is retired.

Include a deliberate change in the sample process: replace an obsolete instruction with a new approved one. Check whether the next draft follows the current instruction. Decide who may approve remembered context before letting a recurring task depend on it. Confirm programme access and commercial terms for your organisation before committing to the pilot.

13 October: Google changes the skills rollout timetable

Our earlier roundup covered the skills announcement. This week’s development is a material availability update: Google says skills in the Gemini app began a gradual rollout on 13 October, expected to finish by mid-November. The separate Workspace rollout is delayed until further notice. Sharing Gemini app skills and Drive integration are described as coming later.

For an IT manager preparing internal training, the app and Workspace dates therefore need separate entries in the plan. Avoid advertising an all-staff Workspace session against the old timetable. Check the linked update before committing to a replacement date.

Our suggestion is to keep the training example useful while waiting. Write a sample task with a known satisfactory result, using invented business information. Ask a colleague to perform it from the written instructions alone. Missing context in that exercise is worth fixing before turning the instructions into a reusable AI skill.

If a pilot user gains access first, record exactly which interface and account were used. A successful demonstration in one place should not be used as evidence that the same capability has reached the rest of the business.

15 October: Anthropic opens deeper access to external evaluators

Anthropic says it has started giving a METR team access broadly comparable to its own risk-assessment staff. It says METR will publish assessments, while some categories of information remain excluded. The announcement was updated on 16 October to clarify the terms.

The distinction matters: Anthropic says it cannot veto publication, but can require redactions within predefined categories and ultimately decides whether information falls into those categories. This is a description of an evaluation arrangement, not a clean bill of health for every Claude deployment.

For a fictional business reviewing an AI supplier, ask which published assessment is relevant to the proposed use. Read what the evaluator could inspect and what remained outside scope. Record the unanswered question that would change your decision, rather than filing the existence of an evaluator as sufficient assurance.

Keep your own workflow test separate. An external assessment of frontier-model risk will not tell you whether a draft response in your customer-service process contains an incorrect delivery promise. Use representative examples and require an accountable person to decide what may reach customers.

A useful outcome for this week

Choose the development that affects an actual plan. If you are evaluating Copilot Studio, produce a small permissions-and-stop test. If you are preparing Google training, correct the availability assumptions. If you are reviewing an AI supplier, record the assessment’s limits alongside its findings.

Leave the review with a named owner and a concrete decision that can be revisited. Deferring a rollout because a required capability is not available is a useful result too.