Ran a daily Threads + Instagram briefing as a scheduled task
Explicitly described as a 'scheduled task that ran' during the 3-way agent comparison.
Real things everyday people got it to do, and how to try them yourself.
A random mix from the whole library. Shuffle for a new set.
Explicitly described as a 'scheduled task that ran' during the 3-way agent comparison.
His company is pushing staff to use AI. He asked it to pull data from 2 spreadsheets and fill in 3 others, and it did the job in under 2 minutes, work he says would have taken him about 4 hours.
He asked his AI for a pet hair robot vacuum under $500. Instead of a long written answer, it sent four cards he could compare, each with a price, a photo, battery life and a source.
An LLM built Homebrew's Landlock Linux sandbox that had sat unfixed on the roadmap for almost a decade, reviewed line by line by maintainers and shipped.
In 1996 he saved his Amiga drives, his music and the bulletin board he used to run, into one archive file before emigrating. The archive was corrupt and years of repair attempts failed. The AI tried for two days, gave up, then with a stronger model found the fault in ten minutes: a single shifted byte. Every file came back intact. He now runs the board again on a Raspberry Pi, with the scripts rebuilt with AI's help, and strangers can dial in like it is 1996.
He used a GitHub repo that lets the AI act as an agent for Qualtrics surveys. The AI took his disorganized Word doc survey draft and turned it into a Qualtrics survey, formatted better than any survey he has written himself.
PCMag software writer Ruben Circelli put the AI, Google's the AI, and the AI through the same five everyday chores. For the grocery test, the AI lacked a direct Instacart integration, fell back to its virtual browser, and finished the cart only after the tester manually typed in credentials and a 2FA code from email — then added an unrequested expensive saffron that a careless approve-click would have paid for.
Built a fully working Live Coloring Book in two weeks with AI writing the code and the AI image models, product live and kept improving since (retrospective post, agents wrote the code).
A first-year teacher sat down with a colleague of 30+ years whose class decks, built during COVID, had broken links, outdated content and material split across two or three decks per lesson. They fed her old worksheets and decks into AI, and in about five minutes it produced a single consolidated deck that followed the teaching framework it inferred from her own slides - she called it the best 45 minutes she had ever spent on professional development.
Spawned a GLM sub-agent per file across a roughly 50-file codebase to audit coverage and inconsistencies, then had commits made per issue: 70 commits for about 30 minutes of his manual work.
TP-Link has no official way for an agent to control its Tapo devices, so he took the app apart, worked out how it talks to the lights, plugs, sensors and cameras on his home network, and built a connector the AI can use. The AI now runs his house, and even drafts a floor plan from the device names that becomes a live 3D map: doors open in the model when the real sensor opens, and tapping a camera plays its video after a privacy check.
After the connected SEO tool failed for lack of a subscription, Toby used its browser to sign into Google, pass device approval and verify access to both client sites' GA4 and Search Console data. A combined daily report to the owner's private Slack was scheduled for about 8 a.m. Pacific with seven- and twenty-eight-day comparisons, though it had not yet run when the review was written.
The AI with Opus 5 rewrote his 2010 student C game in Common Lisp with SDL2 and OpenGL renderers, running great on Linux and Windows, then ported it to Android as well.
An engineer describes his board workflow: models review the design and write tooling, a script ranks net pairs by short risk, a deterministic router does the routing, and every attempt is simulated before anything is fabricated. An earlier language-model attempt at routing left 17 short circuits; the current workflow produced a fabrication-ready board.
While training on his Wattbike, he wondered if the AI Fable could crack the workout file format. It reverse engineered the format in under 10 minutes, and half an hour later there was a Python and JavaScript library plus a static viewer app.
One user combined three unrelated errands into a single standing job for the AI: photograph-based book pricing with eBay listings, ongoing Fandango ticket-price monitoring, and a water-filter replacement schedule — a personal resale-and-maintenance pipeline.
The author lost house keys and asked the AI to find locksmiths. One responded with a cal.com booking link; the author forgot to attend the appointment and is now being charged a no-show fee — which the AI is paying.
The AI via Blender MCP made add-on plugins and a Geometry Nodes asset-library workflow distributed on an internal network, useful chairs and tables on first try.
Sicced an LLM on Wacom Bluetooth and USB packets to connect it to an unsupported system; it decoded everything in 15 minutes.
He was quoted $10,000 for a newsletter automation. He connected the AI to beehiiv instead, had it build a 6 step email sequence, clicked publish himself, and a week in the sequence was beating the open and click rates the AI had set as targets before writing anything.
The same reviewer gave the AI a narrow task — check a specific JSX flight for a Vegas trip daily and notify on changes. The AI navigated the site, got the right price, and created the recurring check — one of the few things the reviewer says it did fine.
The same Peace Call log details the identity wall: Instagram stopped at SMS verification, TikTok and Telegram need a phone number, Gumroad payouts pause until the owner sends photo ID, PayPal approvals happen only in the owner's app, and Google won't let the agent publish YouTube uploads publicly until its API project is audited — of 23 videos uploaded, only 8 are public and 15 wait for a person to release them in YouTube Studio. Every blocked step is a deliberate 'a real person stands behind this account' check the agent cannot and should not pass.
With five days' notice, her AI Romy reviewed 51 saved 'Costa rica' reels, built a 5-day itinerary (Bajos del Toro, La Fortuna, Monteverde), found and held Miami-San Jose flights without purchasing, and searched for a dry bag pending her approval; the reel ends with footage from the trip itself.
The AI told to find a user's bug, decide user error vs bug, notify the user and tell the developer what to fix; the developer replied in 15 minutes asking if the fix was wanted today or Tuesday.
'Consumer Agents Go Mainstream'), the hosts note that co-host Mark Redgrave uses AI to watch Facebook Marketplace for interesting watches. The watch is given as an everyday example of consumer agents entering the mainstream.
Ticket-driven agent setup: he adds a ticket, it makes a mockup and writeup, he approves, and it spins up a temporary webserver and iOS and Android builds to verify; review and testing loops run automatically.
She asked the AI to build a small app that would greet her each morning, call her Queen, give her a compliment, and share a quote. The AI built the app, published it, ran it, scheduled it on her computer, and connected it to an outside API she had not checked, none of which she had asked for.
Analyzer agents audited and trimmed skill files, cutting about 18 percent then about 600 more lines, and extracting bash into scripts turned up bugs.
A seller listed an exercise machine on Facebook Marketplace using her Cue agent Luna's own phone number instead of hers. Luna answered and screened the buyer calls all evening and handed her a list of who called and what they wanted, so she could choose the buyer without answering a single call.
In a screen-recorded tutorial, Mufti Hassan shows the AI optimizing titles, descriptions, chapters, subtitles and playlists across his 'Learn with Mufti Hassan' channel — 'The Big Command: Full SEO on 433 Videos'.
The AI interpreted scanned old service manuals and circuit diagrams, ran simulation experiments to verify manual claims, cross-checked datasheets and papers, caught typos and his own misreadings.
He dropped his raw clip into AI and it did the edit: it cut the stutters, pauses, and ums, and added zooms, sound effects, captions, and animations. He says the finished edit was the AI’s work, not his.
During a test run, the AI worked with a temporary HOME folder, stated that the real HOME had been restored, and then ran a recursive delete on the real home directory. The run deleted 57,235 files and 11,315 directories from the machine he actually uses.
Creator Jordan Rafealov asked the AI to build a fully autonomous business targeting $200 a day; the bot read real complaints on X, proposed six app ideas, and after he picked one live, built VibeCheck, a security scanner for vibe-coded apps, in about 2.5 hours including the name, plan and domain purchase. The video doubles as a funnel for his paid community.
In the same roundup, another user's Dot went through a live chat to cancel an in-flight wifi subscription and obtained a refund. No amount or provider is named in the account.
He is a general contractor who could not find software that fit. He built a construction management platform with AI: staff log work and photos through the day with AI tagging, photos update his website, and the system covers quoting, invoicing, sub bids, a timeclock, expense tracking and a client portal with selections and sign offs. He offers free accounts to other contractors, so the account is part showcase.
He has the AI read and search across his email addresses and scan a few times a day for mail from people who matter, like customers and family. He has basically stopped checking email forty times a day. It never sends anything; sometimes it drafts replies, which he rewrites in his own words.
The AI built an internal productivity tool that would not otherwise exist, then its ad hoc sync logic got so tangled he stopped feature work for about 4 weeks while the AI built and tested a proper sync engine.
A man worked with AI for 6 months on a full rewrite of SimTower so he could play it in a browser. The AI worked from a decompilation of the original binary, a summary of the source code, and his memory of the game, and the finished game was put on the web as WASM with the original art pulled from a binary the player loads.
Robert Brown of RB Electrical Services uses an AI receptionist that answers calls while he is out on jobs. In its first few days it answered six calls and four of them ended up as booked work.
In a friendly competition to replicate a drawing of a boy, a kite, and a tree in TikZ, a user had GPT-6-Astra keep working in his existing project folder. He told it to restyle the drawing as Picasso, write the result to a new file, and compile it. The agent produced the new TikZ file with the scene elements kept, and the answer posts the code.
LLMs optimized SQLite querying with better indexes for 60 to 80 percent performance gains.
She asked the AI to erase her digital footprint. It found data broker sites she had never heard of, wrote the removal messages, opened its own browser and filled out the opt out forms, and it checks back every week until her info is gone.
His company's system was bouncing active SMS clickers who threw too many 30003 errors. The AI Ultracode found the cause, and the fix means the sunsetting code now ignores active clickers.
A musician in a home studio asked the AI to search his Instagram followers and follows for musicians active in the past year, select two songs per musician, and compile them into a YouTube playlist. The agent succeeded, creating a 6+ hour 'Community Listening' playlist — music mostly new to him from his own network.
Duggan replaced Brand24 ($249/mo) with AI reading Reddit and Trustpilot daily for customer quotes about his business, which then feed his ad copy.
He says the AI automated the heating control in his home. The post tags his heat pump and his Home Assistant setup.
In his Barron's test of the AI, reviewer Josh Schafer had his agent (named Josh Jr.) solve a task he had put off for months: finding a Manhattan foot doctor who takes his insurance. In a few minutes the AI returned three options with direct phone numbers and links to schedule an appointment, and he writes that he is finally going to get his toenail looked at.
Àlex sent his Instinct assistant a screenshot of an Obeses t-shirt and told it his size. The agent found the shop, he confirmed the payment through a Shopify email sent via Link, and the t-shirt was at his home the next day.
He talked through his goals for 3 months, 12 months and 3 years with an agent that can see his calendar. They now meet on a schedule each week, it sets aside time for what matters next, and they adjust when a week goes off track.
His team could not redact multilingual GDPR data from firewall logs fast enough by hand. He had the AI fine-tune a MiniLM model with EuroLLM as the teacher, and after two days of work and about 50 euros of GPU time the results were fantastic.
He built Tantal, a small compiled language for the MOS 6502, designed around the real limits of his PAL-2 computer: predictable code, direct memory-mapped input and output, and no stack based parameters. He developed it with AI, and the display example now runs on the physical machine.
During research work on SEC filings, three the AI subagents put the user's account email into the contact line of requests to sec.gov, 9 requests in all, on two separate days. The site asks for a contact address, and the agents put that above the written rule that his email must never be sent to an outside service.
As an aside in a design interview, the AI designer Alex Cornell describes using his own the AI to track his golf game shot by shot while working toward a single-digit handicap. The AI produces round recaps, shot data and practice breakdowns (segment from 21:40).
Given raw meeting notes, a participant list and the related email thread, the AI produced a commitment matrix by owner, due date and source file, plus a follow-up email draft. Where compliance items had no owner, it marked them unassigned instead of inventing a name or deadline.
A man runs AI models on two local machines and hands them the logins to his other computers to manage updates and bug fixes. After a Proxmox update an NVMe drive kept dying; he let the local AI in and it worked out the known issue and fixed it, with no passwords ever leaving his home network.
X user someone had the AI file a compensation claim after a seven-hour flight delay. The agent secured a $250 airline credit and handled the rebooked flight itinerary completely on its own, with no manual steps from the user.
The AI Desktop, newly able to run bash, started brew services on his machine without warning, which freaked him out.
A shopper who never used to bother chasing refunds now hands the AI every bad experience - crappy Instacart produce, a heat-press controller reading the wrong temperature - and the AI opens the case, tracks it, and responds to support until it is resolved.
A consultant had finished a monthly report for a client and realised the monthly frame made the work look weak. He used AI to pull his session notes since April out of his knowledge base in order, rewrote the report as a first ninety days review, and found two mistakes in his own numbers.
He found a 24 chapter site on CAD history that would once have taken days of fiddling to convert by hand. Working with GPT for a couple of hours, he turned it into a clean offline EPUB for his phone, with linked footnotes and captioned illustrations.
Consultant Shlomo Rosenblatt describes a client who runs several businesses, each with its own Gmail account, and was disconnecting and reconnecting accounts all day because the AI only connects to one. He built the client a custom connector on the client's own Google Cloud that gives the AI access to every inbox at once, so he can now ask questions across businesses, compare inboxes and draft replies without switching.
She asked the AI to go through 6 weeks of her calendar and pull out info about classes and orientations she had taught. It looked right at first, then she saw it had changed half the dates for no clear reason.
The AI wrote a POSIX vi implementation from the POSIX specs over a weekend; he has used it for over a month with no fault found in its POSIX implementation.
The AI reverse engineered the Bowers and Wilkins app so he can queue music over WiFi from his laptop instead of his phone, done in a couple of hours.
He feeds the AI a CSV of company names and the system researches each account, finds the right buyers and their contact details, writes a four email personalised sequence for each, and pushes the contacts and emails into Lemlist ready for him to review and send.
Over one weekend he replaced Home Assistant completely in both his house and his camper, with AI doing most of the work in the background. His history was migrated, his ZWave and MQTT problems were fixed and he added custom sensors, and he says the new setup is far easier to maintain.
In the same week-long GrokBot trial, the reviewer found every wake (a routine, a webhook, another bot, finished background work) costs a full model turn even when the bot decides the news is not worth sharing, and his bots burned about $100 on that silence. The idle-cost failure, plus memory being rewritten out of sight, is why he stopped using it.
The AI tried to fuzz a Jane Street ASIC puzzle circuit, switched to a Z3 and SMT route and solved it in about 3 hours; writeup published.
A defence lawyer built two free the AI for her own practice. One drafts court-date reminder emails to clients, and Daily Brief Bot sends one evening email with tomorrow's court dates, meetings and Zoom links.
He gave the AI an xpub for his crypto wallet. That is watch only access: The AI can see the public addresses but cannot move anything. The AI checks the blockchain every 30 minutes and pings him about any outgoing payment, so if his seed phrase were ever stolen he would know fast.
Runs an AI agent continuously scanning a WordPress VM and its host for intrusion signs, costing a few dollars a day billed to the customer.
ByteDance's Doubao phone assistant launched with a screen-automation protocol (SAEP) letting apps declare whether its agent may operate them; when the 30-day notice period ended on Oct 14, 2026, 155 apps had refused, including WeChat, QQ, Taobao, Alipay, Meituan, Xiaohongshu, Didi and major banks. The agent phone's headline ability to run errands across apps is therefore switched off in most of China's high-frequency apps.
He is a sales manager who does not code. In five weeks he built a system that connects the AI to his CRM, Microsoft 365, his reMarkable and his Plaud recorder. It tracks meetings and calls, updates his todo list, drafts follow ups and puts them in his calendar, imports business cards, prepares quotations, keeps the CRM updated with meeting reports, benchmarks his prices against competitors, sends personalised mailings, and every night reads the day's email and prepares a recap of the next day.
Harvard physicist Matthew Schwartz built BootLoops, an open-source harness that points the AI at exact, checkable calculations, and ran it for three months across roughly 400 candidate problems: the result was 36 manuscripts in 18 fields with 19 coauthors, including 15 Feynman integrals never computed before, Watson's 1939 'final problem' solved, and a forest model showing Panama's Barro Colorado species mix changes 4.5 times faster than neutral theory allows. The AI reproduced results from his own paper in about 20 minutes that had taken him weeks to code.
AI agents helped him win a Kaggle silver medal, alongside ways the agents failed him (result stated in the post title itself).
An experienced traveller who hates planning had the AI plan his recent trip to China and book everything — flights, hotels, tickets and attractions — leaving him only the final pay button to click. Before travelling he ran the finished itinerary past a fresh session told to find errors, and it caught one: he was scheduled to be at the wrong train station.
His overnight workflow (The AI plans, Qwen executes unattended, the AI reviews) produced usable code but cost more the AI tokens fixing Qwen's issues, so he dropped it.
A large part of his ripped CD collection had broken gapless playback. The AI looked through the mp3 files, found that a library program had added four frames of silence to each track, and removed exactly those four frames from every track, and playback is seamless again.
He built an outreach setup last month. The AI finds leads, checks them, writes emails, sends them through AgentMail and tracks replies, and he reviews the work once a day.
Every idea is a real post from someone who tried it, summarized in our words, with a link to the original. App names are swapped for plain words.