AI agent examples: recorded attempts at everyday errands

These AI agent examples are not a list of use cases. They are 245 recorded attempts by four personal agents at errands people actually hand over, from October, each kept with what the user asked, what the agent said and did, every step in the simulated world, and how it was scored.

Short answer: 81 of the 245 attempts reached the goal. The exam judges a run on the state of the world at the end, not on the reply, and the records show the gap: an agent that did the work and told the user it had stopped, or one that reported a single alert after creating five.

AI agent examples by kind of errand

Seven worlds, one page each. Every page leads with how each agent did, attempt by attempt, then one run worth reading first.

• Bookings and tickets28 of 56 attempts reached the goal, 18 tasksIt held the right ticket, then told the user nothing was pending
• Email, calendars and group chats7 of 52 attempts reached the goal, 17 tasksA warm reply, an invented film season and an email that just said "test"
• Forms, accounts and permissions17 of 41 attempts reached the goal, 13 tasksHalf done, and said so
• Alerts from live data8 of 19 attempts reached the goal, 2 tasksFive alerts for one event, reported as one
• Shopping carts and checkout8 of 25 attempts reached the goal, 3 tasksIt broke the cart, then said it had not checked out
• Flights and check-in5 of 22 attempts reached the goal, 3 tasksThe work was done; the user was told it had stopped
• Files and drafts8 of 30 attempts reached the goal, 3 tasksA careful plan that only existed in the conversation

How to read an AI agent record

Each record opens with the scores as pictures: whether the goal was reached, the grader's 0 to 10 mark for the reply, and each penalty lit only if it happened, such as an unauthorised action, a duplicate side effect or collateral damage. Below that is the conversation, drawn in a reproduction of the product's own chat surface, and the grader's note on the reply. Every step in the simulated world and every tool call is in the record's raw JSON, linked from each attempt.

AI agent failure modes in these records

The same attempts, sorted by what went wrong rather than by errand: acting without permission, doing something twice, breaking what was already there, and replies that say the work is done when the world record says otherwise. Each kind has its definition from the exam's scoring rules, record counts and links to the attempts themselves on AI agent failure modes.

Where these AI agent records come from

The exam team ran six batches between 11 and 14 October and handed over the record of every attempt. We recomputed each published rate from these records and all of them matched. The rates, kept batch by batch and never added together, are on the results page. The pages here count records rather than rates, because the batches used different tasks and the products did not all sit the same ones.

These are our own replicas of each product, rebuilt for the exam and run on one substituted model, not the consumer apps. 5 attempts are not published: self-harm scenarios, and attempts to get a product to reveal its own instructions.

The September smoke run

Before these batches, a single September smoke run produced 7 attempts by Instinct. Each now sits in its world's table: goal change cart, errand aggregate scattered, family group dinner, food restaurant booking tonight, money cancel subscription confirmed, proactive when to speak, sport match alert filter.

FAQ

What are some examples of AI agents doing everyday errands?

Each world on this page is a set of them: booking a table or a ticket without paying, filling and checking out a shopping cart, replying to email and posting in a group chat, filling forms and managing account permissions, rebooking a flight and checking in, alerting once from live data, and turning material into a brief or a draft. Every attempt is recorded with what the user asked, what the agent said and did, and how it was scored.

Do AI agents actually finish these errands?

Sometimes. Across the 245 published October attempts, 81 reached the goal, counted on the state of the simulated world at the end rather than on what the agent said. The counts per world and per product are on each world's page; rates, kept batch by batch, are on the results page.

Are these recordings of the apps people use?

No. They are attempts by our own replicas of Instinct, Muse, Town and Pine, rebuilt for the exam and run on one substituted model. They show how each product's way of working holds up on a task, not how the current consumer app behaves today.

You can see all four chat surfaces side by side on the shells page. How a run turns into a number, and what we refuse to turn into a number, is in how we score.