Copilot Studio tip. In a workflow, select the Agent icon on the Add panel and pick a published agent, or build one inline. Give it instructions plus data from earlier steps. For risky calls, I'd turn on Request human assistance when unsure. It's slower. Wrong is slower still.
9
Microsoft's new Power Platform update puts business apps in Work IQ into public preview. Users ask for, create and update model-driven app records in plain language, and confirm each change. Fewer clicky screens. Clean up your views and forms first, though.
1
Gemini 4 Carbon is out. Google made it available today to paying developers and enterprises via its API, and to AI Ultra subscribers. It's pitched at long agentic work and cybersecurity. Fine. On a long agent run I'd still have a person check every handoff. Not just the end.
26
Power Apps tip. Microsoft's modern Attachments control is in public preview, and unlike the classic one it works outside a form, so drop it on a canvas, set file count, size and type limits, and save with Patch. One catch. Files only stage until saved. I'd make Save obvious.
8
DeepMind says the 96% came from a proof-of-concept on Mercedes-Benz production lines with Apptronik's Apollo 2. Its own number. A pilot. And it's rolling out to robotics partners only, so I'd treat it the way I treat any software pilot: promising, not adopted.
Gemini Robotics 2.5 is here. 🤖
Our latest physical AI is able to learn hundreds of everyday skills - from folding laundry and serving pastries to autonomously sorting car parts on the factory floor.
Here’s a snapshot of its capabilities 🧵
4
OpenAI put Businesses in ChatGPT into beta. Verified businesses can manage their profile, and eligible US ecommerce and online-service firms can build a Business Agent. OpenAI says user chats aren't shared with businesses. The agent's easy. Clean product data is the real work.
8
Australia plans to legislate mandatory safety rules for frontier AI companies, with a new regulator proposed to enforce them, per the Prime Minister's office. Rapid incident reporting and independent evaluation are on the list. Details to come. I'd want both in any AI rollout.
1
1
23
You need a Microsoft 365 Copilot license to create or edit with Copilot. Viewing and publishing don't. Per Microsoft, there's no scheduling, co-authoring or comments yet, and it's still rolling out. So start small. One status page your users actually open.
6
On a SharePoint site, select + New, then HTML page. Describe what you want and attach files for context, and Copilot opens a draft in edit mode. Small fixes? Type on the canvas. Bigger changes go to the Copilot pane. Then Publish.
1
14
SharePoint tip. You can now describe a page to Copilot and get a working HTML page back. Microsoft says it's saved in your Site Pages library, and Publish shows it only to people who already have access. Here's how.
1
13
Google says TPU 8i Max and TPU 8tx make 4 new TPU systems in 1 year, a first. The 8tx is built for confidential inference for regulated enterprises and governments. No ship date yet. I'd still watch it. In GCC government work, where the data sits often decides the deal.
7
OpenAI says Control Room can stop its research model jobs within 35 seconds of a decision to intervene, and it's working to get that to zero. I'd ask the same of any agent you put in production. Who can stop it? And how long does that take?
5
One catch. Microsoft Learn says an admin has to allow the Copilot Studio MCP client in your environment first. After that, I'd open Edit on the tool, check which tools it gets, and try it in a dev environment before it goes anywhere near real customer data.
3
Copilot Studio tip. Want your agent to answer questions straight from Dataverse? Open the agent, go to Tools, select + Add tool, pick Model Context Protocol, then Dataverse MCP Server. Done. Now ask it "how many accounts do I have?" and see what comes back.
2
13
Meta released Muse Ember, its smallest model, for the stuff agents repeat all day: tool calls, routing, extraction.
Meta says it's $0.10 in, $0.50 out per million tokens.
I'm all for a cheap model doing the boring steps. Just don't skip the part where someone checks its work.
9
Anthropic launched Claude Fable 5.2 and Mythos 5.2. It says they write up to 30% faster than 5.1, at $8/$40 per million tokens, and Fable's on Azure as well.
Best part? Anthropic still calls Opus 5.5 the daily driver.
Fair enough. Big jobs to Fable. Review everything.
1
53
Copilot admin tip. Microsoft's Cost Management API in Graph now lets a Global, Billing or AI Administrator list users' Copilot credit requests and approve or reject them, up to 100 per call. It's beta, so not for production yet. I'd agree the approval rules with finance first.
13
DeepSeek says about 60% of Harness's active users have built or installed at least one of its roughly 35,000 third-party plugins. So the adoption lives in plugins. Fine. In an enterprise, that's also the first place I'd check: who reviewed each plugin, and what can it reach?
8
Read the footnotes. This is Microsoft's own October eval and telemetry data, and it says an agent that keeps calling tools after a failure hasn't necessarily finished the task. Fair. So before switching harnesses, I'd run my own test set and count finished tasks, not attempts.
Microsoft says swapping the model barely moved its agent pass rates. Swapping the harness did: up 10.8 points with an OpenAI model and 22.1 with an Anthropic model, on the same tasks.
That's the part of the Microsoft Copilot harness rename worth paying attention to. After a failed tool call, agents on the new harness kept working the problem in 84 to 91% of cases, versus under 7% on the standard harness.
It matches what I've seen running 30+ bots of my own. Swapping models changed results a little. Fixing retries, memory and triggers changed them a lot.
Read the footnotes, though. These are Microsoft internal numbers, and an agent that keeps trying can still fail the task. Test it against your own work.
6
Skip the agent count. Look at the review queue. Coinbase's CEO still checks each change himself, with screenshots and video clips, a few at a time. That's the bit I'd copy. Don't add 10 agents and sort out review later. Decide who approves what first, then scale.
I've changed how I work over the last few weeks.
I now have one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them. Each lead manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members.
Everything I need to decide or review comes into one prioritized queue, a simple local web app. Each product review includes screenshots and a video demo of the feature. I clear the queue in batches, mostly by voice and keyboard shortcuts. The leads also bring back ideas I hadn't considered.
Working from one queue, instead of switching between dozens of chats, has cut a lot of mental overhead. The system keeps learning my preferences and distilling my judgment into a "Brian brain," so its recommendations and first drafts get better before they reach me. The goal is fewer corrections per change.
Every change still goes through the same human review and security checks. It is eliminating tasks, not human judgment.
What an exciting time to push the frontier of productivity, especially as we build the economy for agents (agentic finance). We should all keep pushing in this direction.
7