The Best Personal Agent May Be the One That Does Less

I have been trying Instinct alongside Muse, ChatGPT, and Grok. If you have not met Instinct yet, it is a personal AI assistant you can text or call, with the ability to use a phone and computer to get things done: arrange travel, make reservations, handle errands.

Its rise has been remarkably fast. Reuters reported on September 28, 2026 that it had raised $1 billion at a $10 billion valuation, and dated its founding to 2025. Forbes reported that its operating company was registered in California in April 2026. Those are different milestones, but either way, this is a very young company attracting extraordinary expectations. Things move quickly in the agent era.

I gave Instinct quite a few things to try: booking restaurants, coordinating dinner plans with friends, and helping me prepare for a business trip the following month. I asked about hotels, places I might enjoy visiting around the work schedule, and how to get around. Its capabilities were good. It was free when I tried it, and watching an agent actually carry out tasks was impressive.

While doing this, I also gave Muse, ChatGPT, and Grok similar requests. I wanted to see how their results compared. What struck me was how nearly identical the answers often felt, especially the recommendations and advice from ChatGPT and Instinct. This was my experience across those tasks, rather than a formal benchmark, but the similarity was hard to miss.

One restaurant reservation brought a different question into focus. Instinct successfully booked it, and I had an awkward realization: I would probably have enjoyed the experience more by opening OpenTable, the restaurant reservation service, and clicking the time myself. 😂

The reservation worked. The question was how much value the work had created.

Taken together, these experiences left me with a thought I cannot quite shake: the most useful personal agent may be the one that leaves more of my life to me, while quietly looking after everything I would otherwise forget.

Instinct itself already points in this direction. Its product description includes following up on dropped threads, alongside arranging rides and booking services. So the interesting question extends beyond one product: what should a personal agent be trying to optimize?

Where I go and where I stay are still my choices

The usual demo begins with a task: book a restaurant, plan a trip, buy something. The agent completes it, and we count a success. But task completion and value creation are different things.

Booking hotels, arranging travel, and shopping are high-value services. There is real money involved, and good assistance can make a real difference. But the value of the service does not tell us how much of the experience we want to give away.

For a business trip, I may simply want a sensible flight and a hotel near the meeting. For a holiday, choosing can be part of the holiday itself. Would I rather stay in a lively neighborhood or wake up to a quiet view? Take an early flight and gain an afternoon, or start the day slowly? Looking through the possibilities helps me imagine the trip. I am doing more than trying to spend less.

Shopping has some of the same pleasure: discovering something, comparing it with something else, figuring out what I actually like. The purchase is only the last step. An agent could make this exploration richer by finding possibilities I missed and explaining the tradeoffs. But “I bought it for you” is not always a better outcome than “Here is something you might enjoy looking at.”

Human effort is not automatically waste. A system that removes all participation may be removing the part I wanted.

Of course, someone else may hate choosing restaurants. Someone with limited mobility, an inaccessible website, or a complicated booking may gain enormously from an agent doing the whole job. The boundary is personal. That is precisely why a personal agent needs to learn which effort I welcome and which effort I want gone.

My shorthand is: automate chores, not joys. Optimize obligations; augment interests.

Payment makes the distinction sharper. With a human assistant, I am quite comfortable saying: prepare everything, then bring it to me to approve. I do not regard the final approval as an embarrassing failure of delegation. It is how delegation works.

An agent can compare options, fill out forms, check terms, and prepare a purchase. I still want the last decision. The few seconds saved by skipping it can be overwhelmed by the loss and recovery work from one wrong purchase. For a familiar, low-risk routine, I might grant a narrow standing permission. That is a deliberate choice about authority, not an assumption hidden inside “help me.”

Delegate the work, keep the authority. Prepare decisions, don't steal them.

And whose decision is being prepared? With my own human assistant, I would not tolerate a financial interest in steering me toward a particular hotel or merchant. If I am trusting someone to help me choose, I need their judgment to serve my interests.

That rules out an appealing business model for the kind of personal agent I want: “The assistant is free; merchants pay it commissions.” Free to me does not resolve the conflict. If the assistant earns money when I book one hotel rather than another, or when I buy rather than walk away, its commercial interests can pull against the job I hired it to do. Disclosing the commission tells me about that conflict; it does not remove it.

I can use a shopping site knowing it sells things. But a personal assistant has a different relationship with me. It knows my preferences, my circumstances, and perhaps my doubts. I do not want that intimacy turned into a more effective sales channel. This is my requirement for a personal agent, not a claim about Instinct's business model: represent me, never the merchant.

A surprisingly useful agent might therefore spend much of its time discouraging transactions. That perk is not worth a purchase you did not want. That upgrade does not solve your problem. Nothing needs doing today. A business measured by transactions generated may find this uncomfortable. Its user may find it refreshing.

The things I keep forgetting

Here is a job I would happily give away: keep checking whether a flight I booked months ago has become cheaper.

I occasionally do this myself. It is boring. The useful result is not a screenshot of a lower fare, either. Is it the same itinerary and cabin? Do the ticket rules allow a change? After fees, lost benefits, or a credit I may never use, is there anything worth bringing to my attention? Only then should the agent ask whether I want to proceed.

The same pattern runs through an ordinary financial life. I have points, miles, credit-card perks, and credits to keep track of. There are forgotten subscriptions, and subscriptions I remember but no longer get enough from. A refund is promised, a deposit is pending, a warranty may still cover something, an insurance reimbursement has not arrived. A rebate or price-protection window closes. Unclaimed property, money or assets held for an owner who has not collected them, may need checking and a claim prepared.

None of these makes for a glamorous dinner conversation. That is part of the opportunity.

Every one leaves an open loop: something unresolved that matters later. “Your refund will arrive in 5–7 business days” sounds like the end of a conversation. For an agent, it should be the beginning of a small responsibility. Remember the promise, retain the evidence, check whether the money arrived, and follow up within the permissions I gave it. Waiting on hold or sitting in a service queue can belong to that responsibility too.

The job is finished when the outcome is confirmed, not when a message has been sent.

The paperwork still exists. It no longer has to live in your head.

This starts to look like someone quietly auditing the loose ends of my life. But money is only one example. The harder limit is attention: I cannot keep watching every low-frequency thing that matters to me.

Consider investing. Suppose I saved a reason for passing on a company: its margins were deteriorating and too much of its business depended on one customer. Months later, new reports suggest both conditions have changed. An Investment Thesis Watcher could bring back my original note, show the new evidence and its sources, and say: “Your thesis may need revisiting.”

That is a hypothetical example of maintaining my own reasoning, not a recommendation to buy anything. The agent would work from my portfolio, watchlist, and saved investment theses. It would notice evidence that weakens my thesis as readily as evidence that supports it. I would decide what, if anything, follows. Observe and surface evidence; do not quietly turn that into deciding and trading.

This broader idea is a Personal Attention Agent. Its working rhythm is simple:

Remember what matters → Watch quietly → Detect meaningful change → Restore context → Ask for human judgment → Track the outcome.

“Restore context” matters more than it sounds. A notification saying “Company X reported earnings” gives me another research assignment. Showing what changed relative to a reason I wrote six months ago gives me a decision I can actually consider.

My preferred measure would be unwanted human attention eliminated. Tasks completed and hours saved miss the difference between an enjoyable hour planning a trip and a few irritating minutes checking, again, whether a refund arrived. This measure has to include the cost of false alarms and mistakes. A silent agent that simply misses everything has protected nothing. It needs a record I can inspect when I want to know what it checked and a clear way to tell me when it has lost access or stopped watching.

Still, most successful checks should end quietly. “No action warranted” ought to count as a good result. The best agent may have no personality at all: more like a background daemon, the old computing term for a program that quietly keeps a service running, than a companion eager to chat.

We had part of this idea a long time ago

Some of this sounds familiar because it is. The early internet had a serious ambition to organize services around the customer. Mobile computing brought another chance to connect the pieces. A short trip into the software attic is useful here:

These were different ideas, but they shared an ambition: stop making the customer carry the work between disconnected services.

I do not think those ambitions were foolish or obsolete. Much of the work was painfully expensive: old systems, incompatible records, integrations that took too long to build and even longer to maintain. There were organizational incentives in the way too. Better software alone does not make a company want to honor a refund.

My good experiences with Delta support illustrate what I still want. The person helping me can already see the relevant context. When they make a change, I can see it in my own Delta app almost immediately. I am describing the experience, not claiming to know Delta's internal architecture. What matters is that the conversation reaches the actual booking record, the authoritative state, rather than ending with a separate promise that somebody will update it later.

The phone call itself is not necessarily the problem. A person can be particularly helpful when something unusual has happened. Retelling the entire story is the problem. I would love a Context Courier that, with my permission, brings the relevant booking, previous attempts, and desired outcome to whoever can help. It need not bring my whole inbox, and knowing my situation must not confer permission to spend my money.

Context is not authority.

This is where AI could change the economics. Instead of only operating the old screens, it could help developers understand the old software and build the connections it was always missing: translate records, expose a usable API, test that a change reaches the right system. We might finally afford more of the customer-centered architecture we wanted a quarter-century ago. That is my bet, not a claim that generated code makes integration automatic.

Voice agents and computer-use agents have a valuable role wherever the old interface is the only available door. I would treat them as a compatibility layer that can remain useful for a long time. But if we can also make the underlying systems work together, teaching machines to imitate all our clicking and waiting need not be the final architecture.

A personal agent needs a home that belongs to me

The more personal an agent becomes, the more important it is that its memory belongs to me.

My preference for a quiet hotel, the reason I rejected an investment, the refund someone still owes me, the promise I made to a friend: these are parts of my life. They should be records in my own data, available to an assistant with my permission. They should not exist only as something one AI company happens to remember about me.

This goes beyond exporting a chat history. An open refund claim needs its receipt, correspondence, deadline, current status, and the next step someone is responsible for. A preference needs to be something I can inspect and correct. A rule about spending needs to remain in force when I change assistants. Otherwise, switching products means reconstructing a working life from old conversations.

I should be able to use ChatGPT today, Muse tomorrow, and a specialized watcher for one particular job. The agents can change; the context and unfinished responsibilities should remain. Credentials should stay under my control, with narrowly granted access. The better an assistant understands me, the more useful it should become, without making it harder for me to leave.

This is why I think in terms of a personal computer for agents. A model can reason about a refund when asked. Keeping responsibility for it over the next six weeks also requires somewhere to run, durable records, permission boundaries, and a way to continue after a session ends or an agent is replaced. A chat window alone is not that environment.

At ArcBlock, this is the direction behind ARC, the Agentic Realm Computer. Its architecture separates the runtime from persistent personal data. DID Spaces provide a user-controlled home for that data; AFS, the Agentic File System, provides a common way for applications and agents to access authorized resources. Blocklets are applications that run on ARC. The design puts continuity in the user's environment, rather than making it depend on the memory of one assistant.

Imagine replacing a refund watcher. The receipt and the unresolved claim stay in my space. I revoke the old watcher's access and authorize the new one to the relevant records. The new watcher should be able to pick up the responsibility without asking me to explain the whole story again. That is a concrete experience I want builders to make dependable on this foundation.

An always-available assistant and user-owned data belong together. The assistant needs continuity to be useful; I need control over that continuity to keep it personal. My agent can be replaceable. My memory should not be disposable.

The opportunity is to build a small, dependable responsibility into that environment. These are ideas for builders, not a catalog of finished ArcBlock products.

A Fare Keeper could look after the flight after the exciting part of buying it is over. A Perk Keeper could remember benefits without inventing spending to use them. A Refund Hound could stay with a claim until the refund appears. The same approach could make a useful Subscription Auditor, Warranty Keeper, or Unclaimed Property Agent. A Promise Tracker could take “5–7 business days” seriously even after everyone else has moved on.

An Investment Thesis Watcher would tend to reasons I once considered important. A Context Courier would help the next person understand my problem without making me start over. Each is a narrow responsibility that has to survive beyond a chat session.

And perhaps a Personal Firewall would become just as important: an agent that filters approaches from merchant agents and marketing systems, admitting what fits my intentions and quietly declining the rest. If every business gets an agent trying to capture my attention, I would like one on my side protecting it.

There is real work hidden inside “quietly.” Someone has to keep the service running, handle lost access, preserve the record, and know when to involve me. I would start with one recurring nuisance and make the outcome dependable before adding more. A watcher that needs constant supervision has handed me a new chore.

Instinct got my restaurant reservation right. That is a useful capability, and sometimes it will be exactly what I need. But the agent I most want would also remember the ticket I bought three months ago and the refund I stopped thinking about last week, without turning either into a stream of updates.

Let me choose where to go, where to stay, and how to spend my time. Keep my memories and unfinished business in a place I control. Take care of the things I shouldn't have to think about.