Here’s what’s shaping AI right now - without the noise:

Google is internally testing a conventional Projects feature for the Gemini desktop app. Dedicated workspaces, their own instructions, their own reference files, and every chat inside the project inheriting the same context.

That is the structure ChatGPT and Claude users have had for years. Google has been routing around it with Gems and NotebookLM-linked notebooks, neither of which behaves like an actual project. Gemini Business has had shareable projects for a while, so the consumer app is getting the thing enterprise already had.

Nothing is announced. It was spotted as a working interface in internal builds, with no date and no confirmation, and Google could still shelve it.

The bigger signal: 👉 Gemini closing this gap removes the last practical reason heavy ChatGPT and Claude users give for not switching, and that moves more seats than any benchmark Google publishes this quarter.

Your inbox isn't the problem. Neither are your meetings. It's the hours you spend on work AI could do in minutes.

285 copy-paste prompts: built for email, meetings, admin, and delegation. Each one has a clear role, a specific task, and a defined output.

No setup. No prompt engineering. Just copy, paste, and get your time back. Free when you subscribe to Mindstream.

Get 285 free prompts

Meta launched Muse, and the architecture is more interesting than the demo. Each user gets an isolated virtual machine in Meta's cloud, monitored by a system called Sentinel, where the agent browses sites and fills forms on your behalf. It keeps running after you close the app.

Two decisions stand out for anyone building in this space. When an app has no official integration, Muse writes its own connector rather than waiting for a partnership. And payments run through Stripe's Link, which issues single-use card numbers, so the merchant never sees a real card and Meta never has to hold one.

Distribution is the part competitors cannot copy quickly. Muse ships inside WhatsApp alongside its own app, which puts an agent in front of billions of installed users without asking anyone to download anything. Free to try, then $20 or $100 a month, US only for now.

See how Meta describes it.

The bigger signal: 👉 Meta skipped the integration marketplace problem entirely by having the agent write its own connectors, and any agent startup whose moat was "we integrate with 200 apps" should read that carefully.

OpenAI's DevDay lands September 29, and platform code points at Managed Agents: create agents and environments, manage them, switch specific skills and plugins on or off, and deploy self-hosted. The shape follows what Anthropic already sells.

The history is short and brutal. Custom GPTs arrived in late 2023, Agent Builder followed in October 2025, and Agent Builder is being wound down after November 30 in favor of the Agents SDK and Workspace Agents. Roughly a year from launch to sunset, because models got good enough to handle pipelines without anyone wiring steps together by hand.

The second finding is stranger. OpenAI appears to be building a path for ChatGPT ads to resolve into an agent instead of a link. You would not read a sponsored answer and click through. You would get handed to an agent that walks you through the purchase, running the conversion funnel itself.

The bigger signal: 👉 If an ad becomes an agent, then the unit Meta and Google actually sell, attention routed to a landing page, stops being the thing anyone buys, and OpenAI gets to define what replaces it.

Projects are the most underused feature in every major AI tool. Most people create one, drop 6 files in, write two lines of instructions, and then wonder why the model still asks what their company does.

The problem is that files sitting in a project are available, not read. Here is how to fix that.

Put your instructions in two places, not one. Write the short version in the project instructions field, because that always loads. Then put the long version in a file named something obvious like 00-read-first.md, and reference it directly in the instructions. The instructions field is small and gets summarized. A file can hold the nuance.

Tell it to read before it answers. Add an explicit line to your project instructions: read 00-read-first.md before responding to anything in this project. Without that sentence, models treat uploaded files as a library to search when something looks relevant, which means half your context never gets opened.

Write rules as positives. "Never use bullet points" performs worse than "write in flowing paragraphs." Negative instructions ask the model to hold an idea in mind and then suppress it, which is exactly the setup that makes it slip out. Say what you want instead of what you fear.

Keep one file that changes and one that does not. Split your context. Stable facts, company details, tone, audience, and constraints live in the read-first file. Current state, this month's priorities, open decisions, and live numbers live in a second file you overwrite regularly. Mixing them means you either rewrite everything constantly or work from stale facts.

Here is a starting structure for the read-first file:

Then in the project instructions field itself, one line does the work:

The bigger signal: 👉 Every session you spend re-explaining your context is a tax you chose to keep paying, and the fix takes about 20 minutes once.

A few tools quietly worth exploring:

Claude can be your analyst, editor, and strategist.

But most professionals are using it to fix grammar.

These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

Claim your free prompts

Three separate companies spent this week solving the same problem from different ends: how does an agent get enough context to be useful without you re-supplying it every time. Google's answer is a workspace, OpenAI's is a managed environment, Meta's is a virtual machine that never shuts off.

The one that wins will be the one people forget they are using.

See you next issue.