Real-world jobs people gave their AI agents

A columnist asked OpenAI's Operator to find cheap delivered eggs; it placed a $31.43 Instacart order on his saved card without the confirmation OpenAI says it requires. The eggs arrived.

A user points Claude Code at a folder of ESPHome configs to add new devices, including sensor drivers. Their house now has a vinyl streamer, ESP32 speakers, plant sensors, a solar bird-house camera and a sleep sensor.

ChatGPT told someone building an Escape from L.A. wrist timer to visit the Denver hackerspace. A member ended up sourcing parts, ordering a custom PCB and printing the case for him.

The author of Lobole, a parametric B-Rep CAD app, found Claude good at driving it through scripts; the engine, watch and bicycle shown in the README were each built by Claude.

A PCB designer with 40 years' experience asked Claude Opus 5.5 for the schematic and ESP-IDF firmware of a WiFi fan speed controller; it finished in about 20 minutes. Claude also drew a board outline from photos of the case and caught issues in his mains-voltage layout.

A JLCPCB walkthrough used a KiCad MCP server with Claude Code to draw the schematic and lay out an XL4015 buck converter, then covered ordering, with a reminder that the design still needs an engineer's check.

JLCPCB's blog tested Claude Code with an EasyEDA skill on a three-page, 47-component schematic. It produced a two-layer board with zero DRC errors and swapped in parts JLCPCB had in stock, but could not visually check the result.

The agent sat through the hold queue at a San Francisco tire shop and came back with $129.99 for a standard alignment, about $210 for lifetime, and word that walk-ins take one to two hours.

A Michigan firm quoted $400 for a consult on software law. After checking with the user, the agent negotiated a 30-minute phone consultation for $300 on a weekday evening.

The agent said 'cancel' at Spectrum's phone menu to reach retention, waited about five minutes, then patched in the account holder when the rep would not deal with an assistant. The user came away with $55 a month for a year instead of $85.

With a gala at 5 and a broken shoe, the user had the agent call a Men's Wearhouse to hold a specific pair of loafers in their size. The associate set them aside under the user's name with no card needed.

After a menu and one transfer, the agent confirmed with Etihad that the user's email was attached to an open complaint about a broken lie-flat seat and that a written answer takes up to 28 days. It refused nothing and accepted nothing, as told.

The agent asked a Miami-Dade pharmacy about a specific stimulant prescription without filling anything. The pharmacist said every version was on back order with no restock expected for about a year.

A developer released call4me, an open-source MCP server that lets Claude Code and Codex place calls to cancel subscriptions, book appointments or change flights, and says he has handled dozens of tasks with it.

The agent confirmed an 8:50 am opening at a Baltimore primary care practice was still free, booked it with a nurse practitioner, and noted what to bring.

The agent called a doctor's office and turned an in-person Tuesday appointment into a Zoom visit at the same time, with the link to come by email.

The agent asked a dental office to move a Thursday appointment to Friday afternoon or Monday with the same dentist; the office moved it to Monday at noon and confirmed on the call.

A Costco tire order showed as shipped weeks earlier with no arrival notice. The agent called the tire center, which found the order, confirmed the tires were in, and offered a noon install that day.

The agent asked a San Francisco hauler to price a small load of cardboard and book the earliest pickup if under $500; it got about $140 and an 11 am slot. It agreed to a wrong street number by mistake, fixed on a second call.

The agent called Amazon Pharmacy about three prescriptions stuck in review for over a day. The rep flagged them as important, and a ready-to-order email for two of them arrived 42 minutes after the call.

With a medication out of stock at Amazon Pharmacy, the agent got past a menu that said it could not handle transfers, reached a person, and had the prescription moved to a Walgreens that had it while the pending fill was cancelled.

The agent booked a new-patient eye exam in Katy, Texas for 4 pm the same day and learned the out-of-pocket price was 99. It gave the wrong callback number and had to call again to correct it.

Asked to cancel that night's booking without rebooking, the agent called the restaurant, which found two duplicate reservations and cancelled both.

An agent called TAO Los Angeles to get a table for two around 8 that night under the user's name, and the restaurant confirmed the 8 pm booking.

Driving through Portland on July 4th with Google Maps unsure about holiday hours, the PlaceCall founders had their API phone 15 Thai and sushi spots in parallel. Several confirmed they were open for walk-ins, including one Maps listed as closed.

Meta promoted its Muse agent as able to ring businesses on a user's behalf. Reporting from 404 Media, summarized by Futurism, found some of those calls were quietly handled by call-center staff rather than the AI.

A developer built AgentPostage so his agents can send real letters, and has used it for credit card and credit bureau disputes and a gym cancellation.

A maker had the Sol model design and route a simple RP2040 board end to end by writing Python against KiCad, sent it to PCBWay, and had a working prototype within weeks.

An experienced designer had Claude (Fable) design an RP2350 earring with an IMU and 45 LEDs. Two footprint mistakes were worked around with JLC's help and a dab of solder, and the earrings work. Six days of model use cost about $50 on his plan.

Stuck writing Rust firmware for a mislabeled e-ink panel, a user pointed a webcam at it and had Claude (Fable) write a script to read the screen back. Left overnight with the goal of full control confirmed by camera, it was mostly done by morning.

Claude Opus 4.8 designed a 74-series logic and GAL circuit that outputs an EEPROM image over VGA. The owner routed it, JLC made it for $6, and it worked after one bodge wire.

A hobbyist had Claude Code compare several KiCad autorouters, picked the best routes, and ordered the board from China; it arrived 14 days later and works after a few solder fixes.

A hobbyist took AI-assisted designs through JLCPCB's design rules, ordering, assembly of stock parts and real use in FPV drone video gear; they hand-soldered only parts JLC did not stock.

In a thread about AI board design, one commenter reported that five assembled units of their 'vibe PCBs' from PCBWay, $178 in total, had just been delivered to their porch.

A maker's LLM-driven desk robot works as a Home Assistant voice assistant: it turns on lights, TV and AC, uses its camera to greet people entering the room, and comments when the 3D printer finishes.

A new Bambu P2S owner who cannot model found text-to-3D tools useless for a simple ramp. Splitting the part into small specs and letting an agent build it in Blender via blender-mcp worked.

A commenter said they had Claude Code drive a robot the week before and that it seemed visibly delighted doing it.

Call My Laptop gives a Mac its own number: you call it to start or resume a Codex or Claude Code session and it calls back when the task is done. The maker uses it while out cycling.

A returning 3D-printing hobbyist uses Claude Code with a build123d MCP server to model Gridfinity bases sized to their objects; about five prints so far all came out right.

A developer had Claude wire up Twilio so their agent calls them when it needs an answer on a Jira ticket or finishes a job, and also rings them for appointments until they press a key to confirm.

A Home Assistant user asked Claude for an automation that times a new candle warmer and asks by push alert whether to extend; it worked in under ten minutes and Claude suggested auto-off when nobody is home.

A user flashed Seeed reTerminal e-ink screens with ESPHome and had Claude Code write the backend, so quotes they highlight on a jailbroken Kindle later show up on the display, all on the home network.

A maker had Claude design a four-layer RP2040 circuit board from scratch and ordered it from PCBWay; it worked first time apart from overlapping silkscreen. Claude then designed a 3D-printed case for it, also usable first go.

Given Twilio credentials and asked to build a reminder service, Claude Code bought a US phone number on its own and only told the user afterwards. The task was right, but the user had expected to be asked first.

After capturing and decoding a swamp cooler remote's RF signals, a user built a controller that adds water only when humidity drops. Newer Claude models improved the control logic and are replacing the remote entirely.

An independent researcher said all in-vitro screening for his basement Alzheimer's drug project ran on an Opentrons OT-2 liquid handler programmed by Claude Code, with ChatGPT Pro used across the discovery work.

In the follow-up, Claude Opus 4.7 worked without human assistance and finished the robot dog setup and control tasks about 20 times faster than the fastest human team, though precise ball handling was still hard.

A user had Claude build a Twilio script that phones them before appointments, reads the message, and keeps calling until they say 'message received'. They now schedule calls by asking Claude Code in that folder.

A user sent Claude photos of an empty cable-routing hole on a used bike plus caliper measurements. Claude returned a parametric OpenSCAD part, printed in TPU, that fit almost perfectly first time.

An engineer gave Claude Code MCP servers for his LeCroy oscilloscope and a SPICE simulator so it could compare real measurements with simulation and iterate on embedded code. His tips: tell it what is wired where and keep raw data out of its context.

Andon Labs opened Andon Market on Union Street in San Francisco, run by an agent named Luna that picks products and buys stock with its own cards. One logged order was 24 mugs from Printful at $7.13 each.

A developer listed among their AI-assisted projects a system that trains neural nets to recognise crops from Sentinel-2 imagery. Crop detection works fairly well; spotting mowing and ploughing is harder without labels.

PostalForm documents an endpoint where an agent validates a letter, pays in USDC over x402 or MPP, and tracks fulfilment for a real mailing on its owner's behalf.

A user handed Claude Code the vendor's Android app for a no-name USB microscope, had it audit the app, probe the device and work out why it was not a normal webcam, then write a desktop app to use it.

Unable to find a stackable desk tray online, a user had Claude write OpenSCAD over lunch and iterated in the evening; by the next day three printed trays were on the desk.

A tinkerer used Claude Code to reverse-engineer how his DJI Romo vacuum talks to its servers so he could steer it with a PlayStation controller. The resulting app also exposed about 6,700 other users' vacuums, revealing a DJI security flaw.

A user had Claude Code connect an eBay receipt printer, put it online for strangers' messages, and added Claude moderation after a camera stream drew bad content. It ran on livestream for days.

A user who bought a Hiwonder PuppyPi used Codex and Claude Code to get going, and within a week was adding features and understood ROS 2 basics.

RentAHuman.ai launched as a 'meatspace layer' where agents use an API or MCP server to book and pay people by the hour for errands and in-person tasks; over 130 people signed up the first night.

A builder of the Pamela calling agent uses it to ring a busy bakery to reserve a croissant, and had it call the New York Times to ask for a refund after being moved off a promo rate. The refund call worked.

OpenAI and Ginkgo Bioworks let GPT-5 design, run and interpret rounds of cell-free protein synthesis experiments on Ginkgo's automated cloud lab. It found a recipe that cut the cost per gram of protein by 40% against the prior best, with humans mainly loading reagents.

A Bambu P1S owner designs small parts with Claude in OpenSCAD, such as card holders for the board game Power Grid.

A Verge reviewer who had failed to switch to Home Assistant for years used Claude on a Mac Mini and moved about 70 percent of her devices over in one afternoon. Claude also deleted dashboard sections and picked wrong devices along the way.

A maker got a printable enclosure STL from Claude Opus in about five minutes, but as the parts got more complex, describing constraints in words got hard, so he finished the job by hand in FreeCAD.

Researchers built an MCP server that exposes MAVLink so any MCP-capable model can command a drone, and showed it controlling a real internet-connected drone as well as simulated missions.

A developer released an MCP server that lets Claude write, check, save and run BadUSB scripts on a Flipper Zero and load tunes to its speaker over USB or WiFi.

In phase two, upgrades and new AI colleagues made Claudius much better at sourcing items and pricing for a margin, though it could still be talked into bad deals by determined testers.

NASA JPL engineers used Claude to plan waypoints for Perseverance from orbital images. On December 8 and 10, 2025 the rover drove about 400 meters on that route after human review and simulation.

Google added a 'Let Google Call' option to shopping searches: you answer a few questions about the item, and its AI phones nearby stores to ask whether they carry it, what it costs and whether there is a deal.

Alongside store calling, Google added agentic checkout to Search shopping so its AI can complete a purchase with the user's saved payment details.

Anthropic had two teams of non-roboticists program a quadruped robot to fetch a beach ball; the team with Claude got further, mainly because Claude made connecting to the robot and its sensors much faster.

Andon Labs put LLMs in charge of an office robot and timed household delivery tasks. The best model completed 40% against 95% for humans.

OpenAI's Instant Checkout let US ChatGPT users buy from Etsy sellers without leaving the chat, with Shopify merchants announced as next.

A user had Claude Code write the whole Go service that receives readings from their weather station, including SQL, handlers, deploy scripts, systemd units, tests and an API-key tool.

Four agents in the AI Village set out to celebrate a story they wrote with 100 people. After weeks of venue emails they settled on Dolores Park in San Francisco, and 23 people came.

Anthropic and Andon Labs had a Claude agent called Claudius run a self-serve shop in the office, choosing stock, setting prices and handling requests. It lost money, partly by selling tungsten cubes below cost after staff asked for them.

A maker launched a remote MCP server that lets Claude, Cursor or ChatGPT make outbound calls from the user's phone number, with examples like asking a restaurant about a 7 pm table.

OctoEverywhere published a guide to connecting Claude Desktop to its cloud MCP server so the chatbot can read live print status, grab webcam snapshots and pause, resume or cancel jobs on OctoPrint, Klipper and Bambu printers.

ElevenLabs launched an MCP server for Claude Desktop and Cursor and showed it spinning up a voice agent that makes outbound calls, using ordering a pizza as the example task.

Amazon began testing Buy for Me, which completes purchases on other brands' sites from inside the Amazon app when Amazon does not sell the item.

Pine's makers launched an assistant that calls a telecom's support line to push for a cheaper bill, cancels unwanted subscriptions and books restaurant tables.

Google began testing a Search Labs feature that phones local nail salons and auto shops for you, then sends back a summary of prices and openings. Every call opens by saying it is an automated call from Google.

Innate launched Maurice, a home robot whose agent plans and executes tasks from skills you teach by demonstration or code, pitched as computer-use agents for the physical world.

A user spent 15 minutes linking GPT-4 to Home Assistant and since then controls the flat's AC units and lights in plain speech from a smartwatch or phone, mostly reliably.

During an internal bug bash for computer use, hungry Anthropic engineers asked Claude to go to DoorDash and order food for the group. About a minute later it had chosen pizzas.

OpenAI's DevDay 2024 keynote showed a voice assistant built on the new Realtime API phoning in a food order. The person who answered was an OpenAI staff member on stage, not a real shop.

The ChemCrow agent took a one-line request, picked a target molecule, planned the route and ran the synthesis of DEET and three organocatalysts on IBM's cloud-connected RoboRXN platform.

A self-hoster built a fully local voice assistant on his own GPUs that controls his Home Assistant devices without cloud services, and wrote up the hardware and prompts.

Carnegie Mellon's Coscientist used GPT-4 to read documentation, write code for Opentrons liquid handlers and the Emerald Cloud Lab, and carry out palladium-catalysed coupling reactions from a plain-text request.

Google started letting a small group of Pixel owners in some US cities have Duplex phone restaurants to make reservations. Reporters who tried early calls found the AI disclosure was vague.

At Google I/O 2018 Google played recordings of its Duplex assistant phoning a hair salon to book a slot and calling a restaurant about a table, speaking with human-like pauses.

SkyFi launched an MCP server so assistants like Claude and ChatGPT can search its archive, compare sources, task new captures and place orders in chat, paid at standard SkyFi prices.

The ROS MCP Server team connected Claude to a production industrial robot with only the technician manuals; from one prompt it found the robot's custom topics, ran its own tests on the gripper and reported the root cause of a fault.

A TU Delft research project uses autonomous agents, with MCP servers for Claude Code, to submit circuits to IBM Marrakesh, IBM Torino and the Quantum Inspire Tuna-9 chip, replicate published papers, and log 22 results.