Today's thread
As AI takes on more work on people's behalf, the question is shifting from 'can AI do it?' to 'who approves what?'. Microsoft announced controls that give AI agents (AI programs that carry out multi-step tasks on their own for a person) an identity separate from employees and narrow their permissions task by task. Anthropic described how a team where Claude writes about 90% of the code decides, by risk level, which changes a human must review.
The two companies point in the same direction. Hand AI broad work, but start permissions at zero and keep humans approving only what is hard to undo. Google's EmbeddingGemma 2 is a slightly different story, yet it fits the same theme of controllable AI by letting companies run a search component in-house without sending internal data outside.
Microsoft gives 60 local agent types their own identity and permissions
On October 19, Microsoft announced on its security blog new controls that treat in-house AI agents as 'the newest insiders'. It added 35 types to its discovery of local agents running on employee PCs, bringing the total to 60, and named Claude Code, OpenAI Codex and GitHub Copilot as examples. Discovered agents land in the same registry as Microsoft Agent 365, so owners, activity and risk can be seen in one place.
What changes is that agents are no longer treated as an extension of a human account. Microsoft Entra Agent ID (an employee badge for agents, added to Entra, which manages staff accounts) now extends to local agents and previously unknown agents found by security sensors, giving each agent an identity separate from the employee who runs it. Permissions start at zero and are granted only as needed for an approved task, for a limited time.
This means the habit of AI coding tools installed by employees inheriting that employee's full account permissions becomes something to manage. As risk rises, responses are graduated: keep reads but block writes, narrow the data and network the agent can reach, or stop execution altogether. High-impact actions can be required to pass verified human approval.
The post does not state general availability, supported operating systems or required licenses. When it becomes usable under your own contract needs to be checked through Microsoft admin center notices or your account representative.
Claude Code team's 7-layer code review cuts reverts by over 40%
On October 15, Anthropic published how the Claude Code team does code review. Claude writes about 90% of this team's code, and from June to September merged pull requests (PRs, bundles of code changes submitted with a request to merge them) more than doubled. Over the same period, the company says, the share of changes that later had to be reverted fell by more than 40%.
The core idea is seven layers of filtering instead of humans reading every change. AI reviews every pull request first, and a change cannot be merged unless it passes a security review and automated checks (a step where machines run tests before merging). Human approval scales with risk: none for low, one person for medium and two for high, while sensitive paths and CI configuration files always need a person. An agent's approval never counts as a human's.
For teams whose AI output already outpaces human review, the lesson is to decide 'what must a human see' before trying to 'review everything'. Internally, the share of PRs passing after one review-and-fix round rose from 65% to 78%, and those needing five or more rounds fell from 8.9% to 1.5%. The productized version, Code Review for Claude Code, is generally available on Team and Enterprise plans, and the /code-review command runs a lighter version on your own machine.
All figures were measured by Anthropic on its own team, with no outside verification. Whether the same effect holds for small teams or teams without solid automated checks is still unknown.
Google EmbeddingGemma 2: multimodal search that runs on a phone
EmbeddingGemma 2, released by Google on October 6, is an embedding model that places text, code, images, audio and video in a single coordinate space. Embedding turns material into 'coordinates of meaning', so it can find related content even when the wording differs. The model has 740 million parameters (the number of numeric settings tuned during training) and is released under the Apache 2.0 license, which allows commercial use.
It is modular, so you load only what you need. The text-only part has 270 million parameters and, in Google's measurement on a Pixel 11 Pro, uses about 191MB of memory; with image and audio added it uses about 567MB. It reads about 8K tokens at a time (tokens are the small pieces AI cuts text into), and output vectors can be shortened from 768 to 512, 256 or 128 dimensions to save storage.
This means internal documents, meeting recordings and product photos can be indexed for search on a company server or laptop without sending them to an outside AI service. It can be downloaded from Hugging Face and Kaggle, making it a reasonable first step to test for a chatbot that answers from internal material.
Performance figures are Google's own, and Korean search quality was not separately confirmed in official material. Until you compare it on your own documents, it is not a strong basis for replacing your current search.
Try today
List the AI coding and automation tools installed on your team's PCs, and add one line for each on whose account permissions it uses and what it reads and writes. Start with the examples Microsoft named: Claude Code, Codex and GitHub Copilot.
Pick just three areas where a human must always check AI output. Start with things that are hard to undo if wrong, such as payments, customer data and outward-facing copy, and write down the rule that an AI saying 'checked' does not count as a human check.
If you use Claude Code, run /code-review once before submitting your next change and read what it flags. If you are on a Team or Enterprise plan, ask your admin whether Code Review is enabled.
Pair 20 questions your company often receives with their correct documents, then compare where the right document ranks when searched with EmbeddingGemma 2 versus your current search.