Last Week in AI: 5–11 October 2026

5–11 October 2026: changes worth putting through a real work test

This week’s selection includes a released small model, Microsoft’s plans for local coding agents and Google’s announcement of a broader workplace agent. For a UK IT manager, the useful distinction is between something you can evaluate now and an announced capability whose access and operating arrangements still need confirming.

We have read the primary announcements, not run comparative product tests. The suggested exercises below are our editorial recommendations. Keep business information within your organisation’s approved arrangements, and use a representative task with a result someone can check.

7 October: Haiku 5.5 offers a new option for narrowly scoped work

Anthropic released Claude Haiku 5.5 on 7 October, positioning it for high-volume tasks such as summaries and classification. Its announcement lists availability on the Claude Platform, AWS, Google Cloud and Microsoft Azure, and introduces an adjustable effort setting for this Haiku generation. Those are product facts and vendor positioning, not a promise that a smaller model will handle every workflow.

For a fictional support team, a useful evaluation would be classifying sanitised tickets into an agreed set of categories. Prepare examples with known answers, including ambiguous cases that should go to a person. Compare the errors and review effort with your current process before allowing any result to route real work.

Anthropic points developers to migration guidance. Treat a model change as a controlled integration change: keep the old configuration available, record the model and settings used, and agree what failure would stop the trial. Do not convert a vendor’s headline cost comparison into an assumed saving for your own workload.

7 October: local GitHub Copilot needs separate checks for inference and actions

Microsoft’s announcement describes local models and sandboxed tools rolling out, while targeting automatic local/cloud orchestration for the end of October. It highlights NVIDIA RTX Spark Windows PCs for the new local coding experience. This is not a claim that the same experience is already available on every ordinary business laptop.

The particularly useful distinction in Microsoft’s explanation is that local inference does not make a session offline. It also distinguishes operating-system containment for child processes from checks applied inside the agent to built-in file tools and remote MCP connections. Avoid treating the word sandbox as a complete description of every tool’s boundary.

Our suggested evaluation is a disposable repository with an agreed output and no production credentials. Before running it, have the administrator document the permitted files and network access. Observe which actions the configuration allows and blocks. If you need an offline workflow, verify that requirement explicitly rather than inferring it from the model’s location. Confirm supported hardware and feature availability before making a purchasing recommendation.

8 October: Google’s Gemini agent announcement makes ownership a concrete question

Google Cloud announced a Gemini agent designed for work across business systems. It describes cloud execution that continues after a laptop closes, and coworker agents with their own identities and Workspace accounts. We are treating this as an announcement: confirm the specific access, commercial terms and deployment availability for your organisation before planning a rollout.

Google describes per-project spending caps that pause an agent when triggered, alongside identity and permission controls. Those stated controls deserve a place in an evaluation plan; they are not an independent assurance that an agent will make the right business decision.

For a fictional operations team, start with a draft weekly status note using an approved sample folder. Name the person who checks the draft and require a separate decision before sending it. Ask how the team would disable the agent, inspect its actions and handle an unfinished task when a spending limit is reached. Keep that ownership clear even if the interface makes the agent look like another colleague.

Leave the evaluation with a decision

Choose the development that matches a service you already support. Write a short acceptance note describing the task, the result expected and the person empowered to stop the trial. Include an awkward example rather than selecting only the easiest input.

Keep the outcome useful: proceed within the tested scope, investigate a specific failure, or defer because access or terms remain unclear. An announcement can justify an investigation without justifying a purchase or an unattended production workflow.