Who Owns the Operating Layer
Spark is Google’s new personal agent. It reads your Gmail, manages your calendar, drafts in your docs, works across your Drive. Announced this week, rolling out next.
Most writeups called it Google entering the AI agent race. Not sure that’s the story, at least from where I’m sitting. The real story is what an agent like this becomes once it’s the thing standing between you and everything you do.
Start with the distinction, because the whole thing turns on it. The agent is just the worker, the part that reads the email and books the table. Plenty of companies can build a competent one. The valuable part is the layer that worker becomes once it holds your memory, your permissions, and a record of everything it’s done for you. The agent is replaceable. The layer is not. Whoever owns that layer owns the most valuable spot in software, because every decision you hand off runs through them. And the reach goes a lot further than your inbox.
Call it the operating layer. A note before the map: this isn’t a complete list. I’m giving you a snapshot, not the full catalog.
Working now. Microsoft Copilot drafts and summarizes across Word, Outlook, and Teams, though reviewers say it won’t run on its own without prompting. Its longer-running version, Copilot Cowork, went live for enterprise in March, built on Anthropic’s models. ChatGPT remembers your chats and files and runs multi-step tasks. Codex, which people still think is just for coding, now reaches into Gmail, Salesforce, Slack, and your signed-in browser, with two million weekly users. Claude ships its own agent, Claude Cowork, and connects out to Gmail, Calendar, Drive, Salesforce, and more. I use many of these daily, both by direct command and running on schedules in the background.
Coming in months. Spark starts rolling out next week to US subscribers on Google’s new $100 plan. It works inside Google’s apps right away. The part that reaches outside Google, into OpenTable, Instacart, Lyft, lands “later this summer,” no firm date. The agent ships next week. The reach ships whenever.
Promised, not working. Apple’s personal Siri. Demoed in 2024, pulled in 2025, rebuilt, now expected this spring. The detail most people missed: Apple pays Google about $1 billion a year to run Gemini inside it. The company with the best privacy story is renting its agent from a competitor.
The standard read might say Google and Microsoft win because they own the apps. The distribution is real. Gemini went from 400 million monthly users a year ago to 900 million last month. But owning the apps isn’t the same as owning the layer that runs on top of them. Google has Gmail, and the agent talking to Gmail doesn’t have to be Google’s. Microsoft has Outlook, same story. A protocol Anthropic wrote, called MCP, lets any agent plug into any app you permit, and Spark will speak it too. Microsoft’s own agent runs on Anthropic’s models. So Anthropic is three things at once: a competitor with its own agents, the engine inside a rival’s, and the standard the others are adopting.
What about everyone else?
Once the layer sits between you and your decisions, it sits between businesses and their customers too. Both sides of the economy run through it.
On the consumer side: ask the agent for running shoes. It shows three options. You pick the first. Options four through forty didn’t lose. They never appeared. You never knew they existed.
On the business side it’s similar. A company’s procurement agent asks for a vendor, a software tool, a parts source. The agent surfaces a short list. Whoever isn’t on it doesn’t get the meeting, the RFP, the renewal. The salesperson who used to get a foot in the door never gets the call, because no human placed the call.
This isn’t theory and it didn’t start this week. Google launched its commerce protocol, UCP, in January, so agents can buy across merchants. ChatGPT already does in-chat shopping. The layer decides which businesses a customer sees, consumer or corporate, and most companies don’t know they’re being sorted by it.
We’ve seen a weaker version. Page two of Google has been “doesn’t exist” for twenty years. Grocery brands have paid for eye-level shelf space for decades. An intermediary controlling what gets seen is old. What’s new is the collapse. Page two still got some clicks. A shopper still reached past eye level sometimes. The option an agent doesn’t surface gets nothing. The intermediary stopped influencing the choice and started making it.
Which raises the question. When the agent picks, is the pick honest?
Right now, by its own policy, OpenAI says yes. Results in ChatGPT are organic and unsponsored, ranked on relevance, and the 4% fee a merchant pays on a completed sale doesn’t buy placement. Take that at face value, for now. But hold two things next to it. The fee exists, and every ad-funded platform in history started organic and ended at auction. And Google launched its commerce protocol in January with an ad product, Direct Offers, running right beside it. Same stage, same week. OpenAI says unsponsored. Google ships the ad slot alongside. Two companies, two instincts, and you control neither.
Even “ranked on relevance” hides a choice. Somebody decides what relevance means, and that somebody owns the layer. No kickback needed for the result to be shaped by whoever built the thing doing the picking.
So, the real questions aren’t about features. Ask yourself these.
Who’s closest to you when you say what you want. Google and Apple have it in your phone. Microsoft has it in your work. Anthropic and OpenAI own no apps and show up anyway, by connecting to whatever you already use.
Who remembers you well enough to be hard to leave. They all keep memory now. None of it is portable. Whoever lets you carry your memory out changes every vendor conversation overnight.
Who you’d trust to act. Booking dinner is one thing. Moving company money is another. Nobody’s shipped a permission system I’d trust at scale. Not Google, whose Spark beta first said it “may share your info or make purchases without asking” before the walk-back. Not OpenAI or Anthropic, whose agents act through your accounts but can’t yet prove to an auditor what they did and why.
Here’s the part that should make everyone stop and think, especially if you run a business. Google has more distribution than anyone and is seen as behind in the contest that defines this year, the agents that do the work. Meanwhile the two companies with no apps to lean on, OpenAI and Anthropic, are setting the pace on the agents themselves, and Anthropic is quietly supplying the protocol the others build on. Distribution was supposed to be the moat. From where I sit, it’s looking more like a starting position.
For everyone else, consumer brand or B2B supplier, it’s simpler. When the agent picks, are you in the set it’s choosing from. And if you are, did you get there honestly, or did the house decide.
I call this the Personal Operating Layer in my book. The layer between what you want and what gets done. Not the agent that does the work, the layer that owns your memory, your permissions, and the record. The agent is the part everyone can see. The layer is the part that’s hard to leave. Spark shipped a version of it this week. Microsoft and Anthropic shipped versions for work. Claude, ChatGPT, and Codex are shipping it now. The Siri Apple promised still hasn’t.
The conventional read says distribution wins. I don’t think it does. The layer that decides what you see, and whether it’s being straight with you, is the thing most individuals and organizations don’t have on their map yet.
Glossary
Agent = the worker. Spark, Copilot, the procurement agent. Always the thing that does the task.
Layer = the position that worker occupies once it has your memory, permissions, and history. The chokepoint. The moat.
About the book
Harry Glorikian is the author of The Invisible Interface: How AI Turns Intentions Into Actions, And Who Wins (Ideapress Publishing, distributed by Simon & Schuster, June 30, 2026). Learn more or pre-order on Amazon.