Hey, y’all. Sherveen here -- been a bit since I’ve emailed, hope everyone is swell.
Through late 2025 and into 2026, I started to talk about the oncoming paradigm of AI agents taking delegated action on behalf of their users.
In fact, I even gave the paradigm a name -- the principal-operator framework -- that I still think the industry should adopt to help us design better experiences as both users and sometimes victims of this new world.
And since then… the trend has exploded in the form of dedicated ‘AI assistant’ products that sit both inside and outside of the main chat apps and coding agents we typically use today.
OpenClaw, Hermes Agent, and Poke (by Interaction) were the first products of this moment, and they’ve since been joined by Grok Bot, Meta’s Muse, the buzzy VC-backed startup Instinct, amongst scores of others begging for attention.
Instinct announced a raise of $1 billion this morning, by the way, despite launching just last month and still being free to users. Beyond that, recently…
Amazon blocks Muse agents -- of course, because Amazon wants you scrolling their interface, seeing their ads, and adding random items to your cart
Meta is literally releasing a Tamagotchi-like device for their agent that’ll let you talk to their cute avatars without having to open your phone
Someone’s Muse agent even invited a buyer from Facebook Marketplace to a guy’s house without warning the user, and was like… ‘hey, the buyer’s downstairs’
People are using their personal assistants to book them reservations and show tickets, to buy items online with a saved credit card, and to delegate digital tasks for work and finance in an asynchronous way.
And OpenAI is rumored to be announcing their own version of a persistent personal assistant during their DevDay event happening tomorrow.
But if you’re like me, you might not have adopted any of these specific products in a serious way.
Maybe you’re exactly like me and you’ve tried them all, and you still use AI agents every day, but the “assistant as text message” form factor hasn’t hit for you yet. Maybe you’re the opposite of me, and your AI adoption is still pretty low, and this is just another in the pile of… “I don’t know what I’d use that for.”
As this trend continues to go exponential, and regardless of whether or not any of these specific products winds up outlasting its competition, I wanted to talk about the 3 things these products are trying to figure out.
If these bots are imagination-agitating about the future, what future are they pointing at? Let’s do a quick tour today so you know what to pay attention to about these products, and why these companies are racing for the pot at the end of the rainbow.
Meta-delegation and umbrella agents
All of these assistants are attempting to wrestle with the same user experience challenge: how do you move away from a chat-by-chat, “start a new thread” mental model of AI to more persistent chats with long-running agents?
The premise comes from both pain and promise: the pain that a lot of users feel, not knowing when to start a fresh conversation or which model number to talk to, and the promise that a “meta” agent can instead be responsible for making those choices for you, dispatching subagents or peer agents on your behalf.
Apps like Instinct want to keep things simple: you text your agent, and everything is handled from that one thread. If a task is too big for the agent to handle itself (given it might be talking to you about 3 things at once), it can spin up dedicated subagents who then report back on completion. All you see is your singular text thread.
Apps like Grok Bot take a slightly different approach: you create different Bots for different purposes. Perhaps you have a ‘Reservation Bot’ for all of your travel and a ‘Solve Bugs Bot’ for handling your software triage. Here, you’re the one assembling the team (though they can share/message each other), which gives you greater visibility and control within specific swimlanes.
And then you’ve got Muse from Meta, which has a “Main chat” that can spin up Side chats that your main agent can boss around, though you can also hop in and talk directly to a side agent at any point yourself.
Two factors will likely wind up dictating the meta-delegation dynamics:
Can shared context across your needs be wisely used to enhance any individual activity, since one agent is seeing across tasks and asks?
Can a singular agent be a good primary talker to you, and a good manager of tasks, subagents, and outside systems, all in the background?
Agent as decision-maker
All of these companies are trying to straddle a tough line: how much autonomy do you give an agent by default? The more autonomy, the more utility, but with that utility comes the added risk of your agent inviting someone on Facebook Marketplace to your home without you realizing it.
This is going to feel very different depending on the use case. For example, expect a lot of delegation-heavy, no-approval-needed workflows to come out of these assistants in the realm of software, where developers and builders are tired of chasing individual threads, remembering which agent worked on which project when, and having to push “go next” buttons on code that they don’t read anymore.
But all of these companies will want you to begin to trust their agent, and their particular flavor of guardrails, permissions, approval processes, etc. It’s a version of a moat in a world where many of these products will wind up looking the same.
And so you’re going to see a lot of effort put into helping the combination of you and your agent-bot make the right choices, quickly, in a way that gives you the combo-feeling of (1) that was better than doing it myself and (2) that won’t go terribly wrong.
Meta is racing toward this sense of safety, building a Secure VM that gives every Muse its own virtual computer that it can use on your behalf (and that you can see it use in the interface).
And all of these companies are building integrations with providers like Stripe and Shopify to enable secure payment credential sharing. Because if they can drive your wallet… everything else is downstream of helping you spend your dollar.
Have your bot call my bot
The final thing you’re going to see a lot of experimentation around -- and I suspect the thing that will be hardest for anyone to get right -- is multiplayer.
Multiplayer can come in a few different forms, but none of it has yet worked in the AI age, and so all of it is up for grabs:
Instinct introduced a ‘Trusted Person network,’ kind of reminiscent of “top fives” or whatever from the old era of texting plans, where your agent will be able to coordinate and communicate with any other person (or their agent) in your Trusted Person list.
OpenClaw introduced a ‘multi-user mode’ to let multiple people control a singular agent, particularly useful for teams of humans working on the same project who all want to steer and give feedback as AI does the hard work.
And then there’s the “my agent talks to your agent, whether any of us opted in” dynamic: people are having their agents call customer service (yes, with voice) and email vendors for support, and in that dynamic, it’s an agent talking to an untrusted person or agent on the other side, and we don’t yet have all the rules & etiquette we’ll need to develop around that.
Whoever finds the key unlock here in agent-to-agent (and human-to-agent-to-agent-to-human, and humans-to-agents-to-humans) will be rich. If there’s a virality factor or a network effect, combined with the rapid action and delegation and asynchronous productivity of agents, you’ll see an explosion of user and engagement growth for whichever platform figures it out.
And there’s the broader thing holding a lot of this together: the race here is a lot more compelling to these companies than it may seem.
It isn’t just that they get to win your eyeballs and sell you more tokens. Remember that throughout everything we’ve discussed, the “assistant” paradigm forecasts asynchronous, delegated activity. That’s new.
An analogy: as Netflix grew into the behemoth it is today, at some point in their history, their CEO said that sleep was their biggest competitor. They had run out of hours in the day with which to provide you entertainment.
Think of the paradigm of software today. We can only be served when we’re paying attention (for the most part). Until now, software was competing in the somewhat zero sum combination of our waking hours at work and at home.
But imagine for a moment that these agents are 10x more effective than they are now, and you are assigning portions of your life to your assistant throughout your day, having swarms of agents do things on your behalf.
Suddenly, we’re breaking out of the old mold. Our agent’s attention and interactivity represent hours we’re making up out of the wind and sand. Those are never-before-seen hours to be monetized. Software beyond a human’s active attention.
What a world.
Alright, I’ll see you tomorrow with quick reactions after OpenAI announces whatever they’ve been up to lately. In the meantime, stay frosty.
Sherveen