Research
Finding something out, then acting on it. Below are 12 recorded examples of people doing exactly this with a messaging agent, with what each one shows and where it came from.
12 examples
An engineer says he got tired of paying for APIs and is using Instinct as the backend of his app: every day it scrapes hundreds of surf reports and writes them into his database. His line is that he is never paying for intelligence again.
Why it is notable
A person using a consumer agent as unmetered infrastructure, on the same day another user posted the agent's own account of a daily browser allowance that stops it working in the afternoon. The two together are the piece on what a billion dollars buys: a free product with no published unit, and a customer whose standing job is hundreds of page fetches a day.
Posted by a user by Aaron Via (@viaaaaron): x.com, 2026-09-23. Added 2026-09-23.
An investor says his wife's Instinct logged her into an Instagram account with 24,000 followers that is not hers, and asked whether she wanted to post something on it. He wonders aloud whether the account could have been taken over with the right prompting, says he is giving Instinct's founder the benefit of the doubt that it will be fixed, and calls it alarming. Posted quoting the Gerber-document thread from five hours earlier.
Why it is notable
The second account in one night of one person's Instinct holding something belonging to somebody else, and the only one where the agent offered to act on it: not a document described but a live session on a stranger's account, with a post offered. It is the row this listing records as holding your password for connected accounts, seen from the other end. No published page of Instinct's describes how a session could reach the wrong user, and the company has not answered.
Posted by a user by Nick Dorsey (@Midnight_Captl): x.com, 2026-09-22. Added 2026-09-22.
A user gave his Instinct and his Muse access to one Google Sheet and told each it was the channel to the other. He says they introduced themselves, agreed a protocol and their own limits on what to share, then took a vendor lead list from Instinct's backlog and split it into lanes, Instinct on national registries and rental-software storefronts, Muse on wedding directories and franchise locators: 4,700 leads in two hours, no duplicates. A second project had Muse, with no access to the code, review his live app as a user and file bug reports for Instinct's builder to fix, which he says found four UX bugs Instinct's own tests had missed.
Why it is notable
The first outside account of agents from two companies working one job between them, and the join is a spreadsheet rather than either vendor's protocol. Instinct's agent-to-agent capability, on its own pages, is Instinct to Instinct; Meta's pages say nothing about Muse talking to another agent at all. The counts are his and the thread is the only account.
Posted by a user by Zev (@ZevLapin): x.com, 2026-09-18. Added 2026-09-18.
On the day Instinct announced Files, a user had it build him a personal CRM in ten minutes: his relationships categorised, a search across them, contact cards with the context of each relationship, and a dashboard mapping his network, which he says had taken him hours to set up in other tools.
Why it is notable
The first outside account of Files, posted within an hour of the announcement, and a use the company's five examples did not list. His profile says he works at Merit Systems, which makes OpenInstinct, the open-source assistant listed here separately, and runs an assistant benchmark site; neither is Instinct's company. Sixty views.
Posted by a user by David Pawlan (@DavidPawlan): x.com, 2026-09-18. Added 2026-09-18.
A publisher gave Instinct her fairly strict parameters for a new apartment and says that four days later she had found it, and moves in two weeks later, back to Manhattan.
Why it is notable
Apartment hunting is the first example the company gave three days later when it announced Files, and here it is done before the feature existed, with the outcome stated: a lease, not a shortlist. What the parameters were and what the agent did in the four days are not in the post.
Posted by a user by Tamara Winter (@tamarawinter): x.com, 2026-09-15. Added 2026-09-18.
A developer who keeps his to-dos in Things, which has no official API, told Muse to build the sync. A minute or two later, he says, his agent's to-dos were syncing to Things across his Apple devices.
Why it is notable
A connector the agent wrote for itself, to a product that refuses to publish one, which is the second account here of Muse extending its own reach with code, after Meta's engineer showed the same. The user is a former GitHub engineer, so the judgement of what counts as sync is an informed one.
Posted by a user by Zach Holman (@holman): x.com, 2026-09-14. Added 2026-09-17.
An Apple engineer with flexible dates had Muse run 146 flight price searches, cash and award, across eleven Spanish arrival cities, ten Italian departure cities and seven dates, to find the cheapest round trip.
Why it is notable
The research sighting with a number on the work: 146 searches is the job a person gives up on after ten. Nothing was booked; the finding is the search itself.
Posted by a user by Deep Mehta (@mehtadeep): x.com, 2026-09-14. Added 2026-09-17.
- MuseWhatsApp
A user connected Muse to Instagram and asked it to steer the feed towards companies that give away products and gift cards for sitting through a demo. Muse found them, signed the user up and put the calls on the calendar; the user reports three pairs of AirPods and a $200 gift card so far, with more demos found for the following week.
Why it is notable
It is an entry in a challenge Meta's chief AI officer launched the same night, #musemoneychallenge, which asks whether your Muse can make you $1,000, so it is a user's post inside a company's campaign, and filed as both. What it shows about the product is specific: the agent changed what a feed shows, then booked the things the feed surfaced, which is the Instagram graph doing what no other listing here can reach.
Posted by a user by kanha korgaonkar (@KORGAONKAR): x.com, 2026-09-13. Added 2026-09-13.
An investor named his Muse agent Alfred and gave it five standing jobs: watch X for the investing topics and tickers he follows, transcribe and summarise podcasts, run a daily news check on the names he covers, watch for proposed regulation and litigation touching his watchlist, and write a morning macro summary. Posted in reply to Meta's chief AI officer saying to use Muse for investing.
Why it is notable
Five recurring monitors rather than five one-off tasks, which is the second published account in two days of Muse doing standing research for an investor, after the trade-alpha dashboard. The jobs are all reading and summarising; nothing here buys, books or logs in anywhere, which is the part of the product every account so far says works. The prompt came from Meta itself, and the reply is the first outside record of somebody taking it up.
Posted by a user by SouthernValue (@SouthernValue95): x.com, 2026-09-11. Added 2026-09-11.
- MuseWeb
An equity analyst had Muse read his trade confirmation emails, benchmark every buy and sell against the S&P 500 from the day of the trade, and plot the notional-weighted alpha over time, with a per-trade view. He used to keep the same journal in a spreadsheet and stopped because the trades had to be typed in.
Why it is notable
The job was abandoned by hand and revived by the agent, which is a cleaner argument for the category than any demo of booking a table. It also shows what a mailbox connector is for beyond triage: the inbox as a data source, and a dashboard built from it in the app rather than a reply in a thread.
Posted by a user by Mostly Borrowed Ideas (@borrowed_ideas): x.com, 2026-09-10. Added 2026-09-11.
- InstinctiMessage
A couple appeared on the Jumbotron at the US Open and wanted the clip. She sent Instinct their seat numbers and a selfie at 11:08pm; by 11:20 it had filed a case with the USTA and emailed the ticket office, the company that produces the tournament and Amex's US Open media contact, and when the production company's address bounced it found that company's web form and went in through that instead. Overnight it went through ESPN's 22-minute highlight package frame by frame checking every crowd shot, found nothing, and kept going after the USTA said no three times on the grounds that the footage is not recorded or archived for public distribution. The next afternoon a director at the production company saw the web submission and passed it on, a producer replied that an editor had cut all the couples from that break, and the footage arrived 18 hours after the first message with no follow-up from her at any point.
Why it is notable
It is the most complete account anybody has published of one of these agents doing a job end to end, and none of it is a booking. What it actually did was work an organisation: find the right four parties, route around a bounced address, absorb three refusals without stopping, and keep going overnight on a search that turned out to be pointless. That is the shape of work people hire other people for, and the only reason it is visible here is that the person kept the timestamps.
Posted by a user by Olivia (@oliviaalevine): x.com, 2026-09-04. Added 2026-09-09.
- MuseWhatsApp
A user asked Muse what it could actually do on Facebook Marketplace. It offered to watch a specific listing or a saved search on a schedule, to search by query, price, location and condition, and to negotiate with sellers over Messenger once that connector is turned on. No purchase was made: this is the agent describing its own reach when asked.
Why it is notable
Two limits arrived unprompted, which is what makes it worth recording. The watch runs on scheduled checks rather than instant detection, so it is not a bidding war tool, and it cannot message a seller at all until Messenger is connected. Both are the sort of qualification a marketing page states as a capability without the caveat, and both come from the agent itself rather than from us.
Posted by a user by Loktar (@loktar00): x.com, 2026-09-08. Added 2026-09-09.