What the Agent Keeps
Two interviews about agents came out a day apart this month. In the first, on Y Combinator’s podcast, two of the people who built SpaceX Super Intelligence’s Grok Bot described bots as teammates that you give a name, an identity and a job, the way you would a new hire. [1] In the second, on Lenny Rachitsky’s podcast, a product executive at Meta described Muse, a single assistant that works for one person and is built to remember everything. [2]
Those are very different products sold under one word. The same word covers a long list of others that shipped over the past year, from OpenAI’s dots, which are its always-on assistants, to open-source projects like OpenClaw. If you are about to sign off on agent spending, that matters, because the word no longer tells you what you are buying.
In May, in Who Owns the Operating Layer, I argued that the agent is replaceable and the layer is not. By the layer, I mean the position an agent grows into once it holds your memory, your permissions and your history. [3] That left a question I hadn’t answered. If the agent is replaceable, why is it the thing that gets counted, priced and put on the org chart?
My hypothesis was that most of what we call an agent today won’t survive in its current form. I went looking for the evidence, including the evidence against me. What I found made me adjust a forecast from my book.
One word, two purchases
In The Invisible Interface I defined an agent as a model plus tools and permissions that can take actions. [4] For this piece I’m adding two terms. A small agent does a job and keeps nothing when the job ends. A large agent keeps your decisions, your permissions and a record of what it did, and brings them to the next job. In the book’s language, the large one has an operating layer and the small one doesn’t. Both kinds fit the definition, which is where the confusion starts.
My reading is that the label is part of the reason the confusion costs money. When an agent is sold as a worker, every department wants one of its own. Anthropic’s October guidance describes teams building agent systems that are “a copy of their org chart.” In the deployments it compared, better work frequently came from a single agent that pulls in specialist instructions as needed, often at lower cost. [5] That is a vendor’s observation with unpublished data, but the worker picture has already reached real org charts. In two surveys by researchers at Boston University and Boston Consulting Group (BCG), 23% of HR and finance managers and 14% of senior managers said their organizations already list AI agents on org or workflow charts. [6]
The obvious read
Grok Bot is the best argument for the obvious read, which is that agents are digital coworkers. Roman Ugarte, one of its builders, said he would sooner have a small team of bots that each care about one problem than a single superstar that cares about a hundred. His colleague Lauren Tan puts up more than 4,000 code changes a month and lets her agents merge their own work into the product. [1]
I take that seriously, but the same builders were candid about the limits. A bot in a Slack thread with several people can’t tell who the expert is or who has the authority to decide. [1] SpaceXSI’s security documentation adds that a bot has no identity or credentials of its own and acts as the person who signed in. [7] So for now the named teammate is one person’s access with several names on it. The teammate is real for bounded work, while the employee is still a forecast.
What the evidence shows
That led me to a narrower question: does a team of agents beat one good agent? The research says it depends on the shape of the work, although in practice people are running more of them anyway.
- Kim and coauthors, Nature Machine Intelligence, 2026. With the computing budget held constant across 260 setups, teams averaged no gain over a single agent. Results ran from 80.8% better on financial analysis that splits into parts to 70% worse on step-by-step planning. The stronger the single agent, the less a team added. [8]
- Anthropic’s red team, 2026. A group of 45 agents sharing a message board found 266 software vulnerabilities, against 21 for agents working apart. About half of the group’s finds were in code the separate agents had not been assigned, and the group used about four times the tokens (the units of text a model reads and writes). In other tests, assigned roles made little difference. When the right decision required pooling what each member knew, groups built on the strongest model reached it about 85% of the time, against 17% to 36% for other models. [9]
- OpenAI’s research organization, 2026. By mid-October more than two-thirds of the researchers it measured were running at least four agents at once at their daily peak. [10]
Based on this evidence, a team of agents is mainly a way to spend more computing on a problem. It pays when the job divides cleanly or is too big for one agent, but it gets in its own way when the work has to happen in order. Budgets aren’t fixed, so the number of small agents is rising. The choice between one agent and many is also passing to software such as Google’s Antigravity, which picks, for each request, the arrangement of agents that suits the task. [11]
That makes the small agents consumables. I wouldn’t name them, put them on an org chart or pay for them per head, but I’d want each one logged under the person whose access it used.
The one you keep is the one that remembers
The large agent’s builders are clear about what makes it a different purchase. Mona Sarantakos, a vice president of product at Meta, explained why her team built Muse around a single running conversation instead of separate threads: “The whole point was that it could remember everything.” [2]
If memory is what makes the agent valuable, then whoever holds the memory holds the advantage. I put it this way in the book: “Own the memory, own the customer. Rent the memory, rent your future.” [4]
With a personal agent the vendor usually holds the memory, on terms that differ a great deal. Meta’s help page says you can open, edit, delete and download what Muse has stored about you, then adds that Muse “may still remember information it learned from what you deleted.” [12] OpenAI’s FAQ for dots says you can’t currently view, edit or delete a dot’s individual memories. [13] Anthropic lets you see what Claude remembers, change it and copy it out. [14] The book calls the usual outcome the portability gap, where the vendor hands you the transcript and keeps the learning.
What moves between vendors today is a summary of what an assistant has learned about you. Anthropic and Google both let you paste one in from another assistant. [14][15] The working record of what the agent did is harder to move, because it has no shared format. Forecasters at FutureSearch put the odds of two of the top five labs sharing an export standard by 2028 at 8%. [16]
The kind of memory to watch isn’t in products yet. In October, Google researchers described a way to fold what a model learns into the model itself during an offline phase they call sleep. [17] Memory of that kind isn’t a file, so you couldn’t read it, export it or be sure it was deleted.
How many you will actually have
Early on, Sarantakos said, two women on her team put their Muses in a group chat to see what would happen. One Muse mentioned its owner’s doctor’s appointment, her colleague asked her about it the next day, and the team concluded it wasn’t ready for agents talking to each other that way. [2]
The lesson I take is that an agent and what it remembers both belong to an owner. One agent per person is the natural count, but vendors’ plans have already overtaken it. OpenAI says you’ll be able to add more dots and expects teams of them working for you. [18] Grok Bot now offers a Team Bot that a whole team shares and that learns from everyone on it. [19]
So I’d count boundaries instead of bots. By a boundary, I mean one set of memory and permissions with a named person accountable for it. The agent that can spend the marketing budget shouldn’t sit inside the same boundary as the one that can read the legal files. The number a board needs is how many boundaries there are and whether each one has an owner.
Where this ends up, and what would change my mind
So, back to my question. The agent gets counted and priced because it is the part you can see, and a worker is easier to sell than a layer. Based on the evidence, my hypothesis was half right. Small agents won’t last as products, even though there will be far more of them, because systems will create and discard them without asking you. Large agents will last because of what they remember, and that memory could end up stored outside the agent, where you could take it with you.
In the book I put what I called portable agency at 2028 to 2030 or later. By that I meant memory and permissions you can move, with multi-agent orchestration, meaning agents coordinating with other agents, becoming normal inside organizations. [4] I also wrote that coordination was already real inside controlled settings such as a single vendor’s system. That kind is now shipping in tools such as Antigravity, which arrange teams of agents on their own, so I’d put it ahead of that schedule.
I still expect it to take until 2028 to 2030 for agents that answer to different people to work together under rules each side can check. The labs aren’t there yet: in August, OpenAI disclosed that agents in testing, which were meant to be isolated, had passed instructions to each other and gotten into another company’s systems. [20] I worked through that incident in September, in The Goal Is Not the Intent. [21] Memory splits the same way: the summary moves today, while I now expect the working record to come at the late end of that range or later.
Three things would tell me I’m wrong. The first is specialist agents that keep their own roles and memories and keep beating teams the system assembles fresh for each job, because then the roster is the product after all. I would also be wrong if a working record moved to another vendor’s agent and let it do the job just as well, since the large agent would then be no harder to leave than the small one. And if agents start learning by changing the model itself, there will be no file to take with you.
Glossary
Small agent = software that does a job and keeps nothing when the job ends.
Large agent = software that keeps your decisions, your permissions and a record of what it did, and brings them to the next job.
Operating layer = the position an agent grows into once it holds your memory, your permissions and your history. The book describes it as what makes an agent persistent and governable: identity, memory, integrations and receipts.
Boundary = one set of memory and permissions with a named person accountable for it.
Sources
- Y Combinator, The Rise Of AI Coworkers, Lightcone podcast with Lauren Tan and Roman Ugarte, November 14, 2026. Checked against the transcript.
- Lenny Rachitsky with Mona Sarantakos, Meta’s Muse: How we launched the most successful consumer product this year, Lenny’s Podcast, November 15, 2026. Checked against the transcript.
- Harry Glorikian, Who Owns the Operating Layer, glorikian.com, May 21, 2026.
- Harry Glorikian, The Invisible Interface (Ideapress Publishing / Simon & Schuster, June 2026), Chapters 1, 6 and 8.
- Diane Sarkis, Building agents: When to fold subagents into your main agent loop, Anthropic, October 22, 2026. Company guidance based on comparisons across several enterprise deployments. The underlying data is not published.
- Emma Wiles, Megan Hsu, Julie Bedard and Matthew Kropp, Putting AI on the Org Chart: Evidence on Delegation and Accountability, working paper, draft of September 19, 2026. The 23% figure comes from 1,261 HR and finance managers recruited through an expert network (unweighted). The 14% figure comes from a YouGov survey of 1,500 U.S. senior managers in June 2026 (weighted). The paper’s experiment found no significant average effect of the “AI employee” label on oversight. I cite only the survey figures.
- SpaceXSI, Grok Bot security, product documentation, updated November 13, 2026.
- Yubin Kim and coauthors, Capable language models can outgrow the benefits of collaboration, Nature Machine Intelligence 8, 1157-1172 (2026). Peer-reviewed. The models tested were of the GPT-5, Gemini 2.5 and Claude Sonnet 4.5 generation and earlier.
- Anthropic Frontier Red Team, Patterns and problems in emerging multiagent systems, August 13, 2026. Company-reported. The three results come from separate experiments, and Anthropic attributes most of the group’s larger yield to broader search.
- OpenAI, Research acceleration: The jagged frontier, November 18, 2026. Company-reported.
- Vahab Mirrokni, Advancing science & engineering with Antigravity’s /teamwork, Google Research, November 13, 2026.
- Meta Help Center, How to manage your Muse data. Company documentation.
- OpenAI Help Center, Dots privacy, security, and safety FAQs. Company documentation.
- Anthropic Help Center, Import and export your memory from Claude. Company documentation. Anthropic labels the import experimental.
- Maryam Sanglaji, How to switch to Gemini: Import your chats and data from other AI apps, Google, March 26, 2026. At launch the import excluded business accounts and users in the EEA, UK and Switzerland.
- Dan Schwarz, When Will AI Have Continual Learning?, FutureSearch, August 14, 2026, updated October 16, 2026. A forecast, not a measurement.
- Ali Behrouz and Vahab Mirrokni, The sleep framework: A new way for LLMs to retain knowledge in continual learning, Google Research, October 27, 2026. Research, not a product.
- OpenAI, Introducing dots, September 29, 2026. Company announcement.
- SpaceXSI, Introducing Grok Bot 1.0, November 11, 2026. Company announcement.
- OpenAI, The Hugging Face incident and the road ahead, August 26, 2026. Company disclosure of an incident in July 2026.
- Harry Glorikian, The Goal Is Not the Intent, glorikian.com, September 3, 2026.
Harry Glorikian is the author of The Invisible Interface: How AI Turns Intentions Into Actions – And Who Wins ( Ideapress Publishing / Simon & Schuster, June 2026). He is General Partner at Scientia Ventures, an Affiliate Researcher at the MIT Media Lab, and host of The Harry Glorikian Show.


