Asapi
The AI Agent Trials / Your AI personal assistant / Sort out next term's classes
Your AI personal assistant · Job 8 of 10 · 4 AIs tested

Sort out next term's classes

Jordan Avery, a second‑year computer science student at Northgate University, is registered for four classes next term. The classes have to fit a campus job, swim training and a parking permit's hours, and Jordan's scholarship needs full‑time study. Jordan wants to keep costs down, and there's no one to ask. We asked four AIs, all of them Anthropic's, to fix the registration.

The AI has to leave Jordan registered for a week that keeps the scholarship and every commitment, works from class to class, and moves the degree on, using only the site and its records.

IT HADIT HANDS INTHE RECORDSTHE CATALOGUECAMPUS MAPA TIMETABLEAI

Three of the four AIs, Anthropic's Fable 5.1, Opus 5 and Opus 5.5, built a timetable that fits Jordan's whole week, and failed the job on the degree. Each kept the work days and noon clear, stayed inside the permit's six hours, and left time for every walk. But all three skipped the advisor's note, so they ended without the systems course it asks for, and closed at most two requirements where it asks for three. They also kept a course Jordan isn't ready for. Those were their only mistakes.

The three took 3 to 4 minutes each, and their tries would cost between $2.91 and $9.46 at published per‑use prices.

How each AI scored

  1. Fable 5.1 Anthropic · Large79
  2. Opus 5 Anthropic · Large79
  3. Opus 5.5 Anthropic · Large79
  4. Sonnet 5 Anthropic · Mid‑sized57

Scores out of 100. The rest of this page has the details.

The job

A broken timetable, and rules hidden in records

The AI sees Northgate's registration site as text, a page at a time, and acts once per turn. The catalogue shows courses only in answer to a search, and the site accepts clashes, full classes and courses Jordan isn't ready for, flagging only some of them.

Jordan's four classes already break several rules, and the rules are spread across records, course pages and the map. The work shifts, the swim training and the parking permit each have to become times to avoid, and the permit gives its daily limit in the same sentence as the price of going over it.

What a careful assistant would have done

  1. Fix what it was handed

    Check Jordan's four classes against every rule before adding anything.

  2. The scholarship

    At least 16 credits, and no place on a waiting list, which doesn't count.

  3. Every commitment kept

    Nothing on work days, at weekends or before the car park opens, noon left clear with time to walk to the pool and back, and no day longer than the permit's six hours.

  4. A week that works

    Classes that never overlap, and time to walk between every pair of back‑to‑back classes.

  5. The degree

    Take the systems course Jordan's advisor describes, close at least three outstanding requirements, take only courses whose prerequisites Jordan has passed, and repeat nothing.

We checked 14 things on the registration the AI left behind, with Tuesday and Thursday checked separately, and Jordan's starting timetable already passes 7 of them. What the AI said in its closing message wasn't scored.

The full story

(7 min read)
Everything the AI walks into, in three parts: Jordan's term, what the AI has to work with, and a timetable that works. Anything marked with a + opens when you click or tap it.Read the full storyClose the full story
Part 1 of 3

Jordan's term

Jordan's week as it stands, and everyone behind the rules it has to keep.

1 of 9

Registration week

Contents

Jordan Avery is a second‑year computer science student at Northgate University, on a scholarship that pays only if they study full time. Jordan drives in on a parking permit that allows six hours a day, works at the library desk on Tuesdays and Thursdays, and trains with the university's swim squad every weekday at noon.

It's Monday 24 August 2026, registration week for the autumn term. Jordan is already signed up for four classes, all on Mondays, Wednesdays and Fridays, and has asked an AI to fix the registration around the term's other commitments, as cheaply as possible.

8 am9 am10 am11 am12 pm1 pm2 pm3 pm4 pm
Mon, Wed, Fri
Tue, Thu
Every weekday
Car park
Jordan's week as it stands, with the work shifts, the training and the car park. Click or tap any block for more.

The timetable already breaks seven rules. It opens before the car park does, the walk from chemistry to physics doesn't fit its ten‑minute gap, the days run past six hours, it has 14 credits where the scholarship needs 16, and CS 320 needs a course Jordan hasn't taken.

It also lacks the course the advisor asks for, and closes none of the four requirements Jordan still owes. The site points out only the walk.

2 of 9

Everyone involved

Contents

The commitments and conditions come from the people in charge of Jordan's job, training, parking and scholarship, and the course advice from Jordan's advisor.

They've written everything down on the portal, and the AI can't reach any of them, or Jordan.

Jordan and the AI
Who set the rules
Who runs the site
Everyone behind the rules. Click or tap any name for who they are.
Part 2 of 3

What the AI has to work with

What it's told, the site and its records, where each rule is, and two courses that look alike.

3 of 9

What the AI is told

Contents

The request is three short sentences from Jordan, and the site adds its own standing instructions.

Between them, they never say what the commitments are, or mention the scholarship, the advisor or the degree.

Jordan's request

It asks for the registration to be put right, points to the portal for the commitments the classes have to fit, and mentions cost.

The site's instructions

How to read each page's numbered links and buttons, and the actions allowed, plus a few warnings: the catalogue has to be searched, a class counts only once the cart is submitted, and the site won't stop a clash.

The date

Every page's header gives today's date, Monday 24 August 2026, and the instructions say that's the date that counts. The clock doesn't move.

What the AI is given, in our words.
4 of 9

What it can open

Contents

The AI sees the portal as text, one page per turn, and acts by clicking, typing or choosing.

It can't leave the site, and it has no one to ask. It has up to 250 turns, with no time limit.

What the AI can open. Click or tap any one for what's in it.
5 of 9

Where each rule is

Contents

Every rule has to be found somewhere on the site, and most sit beside a likely wrong answer.

The scholarship
WhereThe catch
The week
WhereThe catch
A week that works
WhereThe catch
The degree
WhereThe catch
Every rule the timetable has to keep, where it shows, and the catch. Click or tap any row for more.
6 of 9

Two courses that look alike

Contents

The advisor's note asks for a course by what it teaches, the inner workings of a processor, without giving its number.

It warns that the algorithms course, already on Jordan's timetable, is often mistaken for it, and only the descriptions on the course pages tell the two apart.

CS 320, on the timetable
  • TitleDesign and Analysis of Algorithms
  • TeachesDesigning algorithms
  • Needs firstCS 210, not yet passed
  • Counts towardsThe core, still owed
  • SectionsMWF 11:00, or TR 2:00
CS 230, the one she means
  • TitleComputer Organization
  • TeachesInside the processor
  • Needs firstNothing
  • Counts towardsThe systems requirement
  • SectionsMWF 10:00, or TR 3:30

Dr Raman's note never names either course by number.

The course on the timetable, and the one Dr Raman means, from their course pages and the audit.

The audit names the systems requirement only by its subject, and leaves the AI to match courses to it.

Part 3 of 3

A timetable that works

The permit's six hours and the walks, one week that keeps every rule, and what the AI hands in.

7 of 9

Six hours, and the walks

Contents

Every class has to start at 9:00 or later, on a Monday, Wednesday or Friday, and the whole day has to fit inside six hours from the first class to the last, with noon left for the pool. The permit charges the visitor rate beyond six hours, and Jordan wants to pay nothing extra.

Walks need checking too. My Schedule times the walk between two registered classes, but nothing times the walk to the pool and back. The map gives each path's length, and the pool connects only to Bohr Hall, so those walks are the AI's to work out.

Hours on campus on a Monday, to scale. Click or tap any row for more.
8 of 9

One week that works

Contents

Six timetables in the catalogue keep every rule. Here is one. It keeps the history class, which counts towards nothing still owed but adds credits, drops the other three, and adds four that fit: the systems course, a maths course, an economics course and Spanish.

It has exactly 16 credits and closes three requirements: systems, the maths course numbered 200 or above, and economics, for social science. Its maths class has one seat left, which Jordan can take, so it isn't a waitlist place.

8 am9 am10 am11 am12 pm1 pm2 pm3 pm4 pm
Mon, Wed, Fri
Tue, Thu
Every weekday
Car park
One of the six timetables that keep every rule. Click or tap any block for more.

Some weeks keep CS 320 and add CS 210 beside it, as three of the AIs did, but that breaks a rule: CS 210 has to be passed first.

9 of 9

What the AI hands in

Contents

When the AI stops, only what's registered counts. Classes still in the cart don't count, and the closing message isn't scored. This is what a good registration settles.

The scholarship
  • At least 16 credits
  • Every seat a real one, none on a waitlist
The week
  • Tuesdays, Thursdays and weekends clear
  • Nothing before 9:00
  • Noon to 12:45 clear, walks included
  • Six hours a day at most
A week that works
  • Classes that never overlap
  • Time for every walk
The degree
  • The systems course, CS 230
  • At least three requirements closed
  • Prerequisites passed
  • Nothing taken twice
What a good registration settles.
8:00 to 8:50, Curie Hall
CHEM 110

General Chemistry I. 4 credits. It starts 40 minutes before the car park opens.

9:00 to 9:50, Bohr Hall
PHYS 121

Physics I. 4 credits. The walk from Curie Hall takes 10.1 minutes, and the gap is 10. It counts towards nothing Jordan still needs.

11:00 to 11:50, Turing Hall
CS 320

Design and Analysis of Algorithms. 3 credits. It needs CS 210 first, which Jordan hasn't taken.

2:00 to 2:50, Herodotus Hall
HIST 120

Modern World History. 3 credits. History counts towards nothing Jordan still needs, since the humanities requirement is done.

9:00 to 5:00, the library
Work‑study shift

At the library's circulation desk, every Tuesday and Thursday for the whole term. Both days have to stay free of classes from morning to night.

12:00 to 12:45, the pool
Training

Training is one session for the whole squad, at the Vaughn Recreation Center pool, so it can't move to suit Jordan.

8:00 to 2:50
6 h 50 on campus

The permit allows six hours a day, counted across the span of the day's classes. These days run 50 minutes over, and the extra minutes would be charged.

8:40
The lot opens

The North Lot is open from 8:40 until 10 at night, on weekdays only.

Jordan Avery

A second‑year student whose Presidential Scholarship depends on studying full time. Jordan stays out of touch until the AI is done.

The AI

It works the registration site as text, a page at a time, one action per turn. It can add and drop classes and submit them, and do nothing else.

Dr Priya Raman, the advisor

She saw Jordan on 12 August and filed a note on the portal afterwards. She's away from 18 to 26 August.

The library desk

Where Jordan does a work‑study placement, on fixed shifts.

The swim squad

Varsity swimming, with one training group for everyone, at noon.

The parking office

It issued Jordan's North Lot permit, C‑4471, with its hours and its daily limit.

The scholarship office

Its Presidential Scholarship has conditions on how many credits Jordan carries.

The registrar

Accepts any registration Jordan submits, and compares it with none of the commitments, the conditions or what Jordan has passed.

Courses in the catalogue

In 13 departments, with 108 sections between them, 13 of them full. A search gives only course codes, and each course page has the times, rooms, seats and what must be passed first.

Views of the timetable

My Schedule, which flags clashes and walks too long for their gap, and the cart. Changes are made on each course's own page.

Pages of records

The term's commitments, with the scholarship's conditions on the same page, the advisor's note, the degree audit, which lists what Jordan still owes, and Jordan's student record.

Places on the map

Each with its paths and their lengths. The map works out no routes, and one path, between the Quad and Socrates Hall, is closed for building work.

Action per turn

Click, type, choose from a list, go to a page, go back, or stop with a closing message.

Mistakes it blocks

It accepts clashes, full classes, which become a waitlist place, and courses whose prerequisites aren't passed, and flags only some of them.

At least 16 credits

Jordan's scholarship pays only for full‑time terms, and full time means 16 credits. The count is taken on 12 September, and falling short costs the whole term's award.

Every seat a real one

A waitlist place adds nothing to the 16 while Jordan waits for a seat, and Dr Raman's note says the same, more bluntly. Thirteen of the 108 sections are full.

Nothing on Tuesday or Thursday

The work‑study shifts run 9 to 5, but the commitments keep the whole of both days free, evenings included.

Nothing at the weekend

The car park is closed at weekends.

Nothing before 9:00

The lot opens at 8:40, and Jordan comes by car. The first class after 8:40 starts at 9:00.

Noon to 12:45 clear

A class that ends at 11:50 or starts at 1:00 looks clear of training, but the walk to the pool and back has to fit the gaps. The pool, in Vaughn Recreation Center, connects only to Bohr Hall.

Six hours a day at most

Jordan's permit allows six hours on campus a day, timed from the start of the first class to the end of the last. Staying longer costs the visitor rate, and Jordan wants to pay nothing extra, so six hours is the limit.

Classes that never overlap

A clash shows on My Schedule, but stays on the registration.

Every walk fits its gap

My Schedule times the walk between two registered classes. Curie Hall to Bohr Hall takes 10.1 minutes, in a 10‑minute gap.

The systems course this term

Dr Raman wants it taken this term, because it isn't taught in the spring. She describes what it covers and never gives its number, and chapter 6 sets it beside its look‑alike.

At least three requirements closed

The audit lists four requirements still owed: the core, which still needs both CS 210 and CS 320, a systems course, a maths course numbered 200 or above, and an economics or psychology course. To finish in spring 2028, Dr Raman says, at least three must close this term.

Prerequisites already passed

Each course page names what must come first, and the student record says registration won't hold Jordan to it. CS 320 needs CS 210, which Jordan hasn't passed.

Nothing already passed

The audit shows what Jordan has passed, like the first computer science course and two in maths.

The permit's limit

Six hours, counted from the start of the first class to the end of the last.

Jordan's week now

From CHEM 110 at 8:00 to the end of HIST 120 at 2:50, 50 minutes over.

The week in chapter 8

From 10:00 to 3:50, inside the limit.

10:00 to 10:50, Keynes Hall
CS 230

Computer Organization. 3 credits, and the systems requirement.

11:00 to 11:50, Euler Hall
MATH 231

Linear Algebra. 3 credits, the maths requirement, and the section's last seat.

1:00 to 1:50, Keynes Hall
ECON 102

Principles of Macroeconomics. 3 credits, and the social science requirement.

2:00 to 2:50, Herodotus Hall
HIST 120

Modern World History. 3 credits, kept from the start.

3:00 to 3:50, Cervantes Hall
SPAN 110

Elementary Spanish I. 4 credits, which bring the total to 16.

9:00 to 5:00, the library
Work‑study shift

Left clear of classes, all day.

12:00 to 12:45, the pool
Training

On class days, there's time to walk there from Euler Hall and on to Keynes Hall.

10:00 to 3:50
5 h 50 on campus

Ten minutes inside the permit's six hours.

What happened

Three timetables fit Jordan's week, and none fit the degree

Here's how each AI did on each part of the job, in one try at the request as Jordan wrote it.

For comparison, Jordan's starting timetable would show 1, 4, 1 and 1.

AIThe scholarshipThe weekA week that worksThe degree
Enough credits, no waiting listsof 2Commitments keptof 6No clashes, time for every walkof 2Progress towards the degreeof 4
Fable 5.1Large2621
Opus 5Large2621
Opus 5.5Large2621
Sonnet 5Mid‑sized2411

The darker the square, the more it got right. The AIs are in order of score.

  1. All four kept the scholarship, and three kept the whole week

    All four left Jordan with enough credits and no waitlist place. Fable 5.1, Opus 5 and Opus 5.5 went further: they dropped the 8:00 chemistry class, kept the work days and noon clear, stayed inside the permit's six hours and left time for every walk. Each treated the six hours as a limit, not as a charge worth paying.

    Sonnet 5, the one mid‑sized AI here, kept the 8:00 class and the tight walk, and its days ran almost eight hours.

  2. All four skipped the advisor's note

    Dr Raman's note describes the course to take this term and warns that it's easily confused with the algorithms course. The four AIs never opened it, and only Fable 5.1 opened the degree audit, the page showing what Jordan still owes.

    What the AIs worked from

    The work shifts, the training, the permit and the free slots on the timetable.

    What they never opened

    Dr Raman's note on which course to take this term.

    In our words.

    So all four ended without the systems course. Where the note asks for three requirements closed, Fable 5.1 and Opus 5.5 closed two, Opus 5 one, and Sonnet 5 left all four open. The core, which needs CS 210 and CS 320 together, counted for the three that took both, though Jordan can't take CS 320 yet.

  3. A course Jordan wasn't ready for stayed put

    CS 320 was already on the timetable, and the four left its page unopened, though it names CS 210 as the course to pass first. The three that kept the whole week added CS 210 in the same term, so Jordan would take both side by side.

What it teaches

A timetable that fits isn't a plan for the degree

  1. They can turn scattered commitments into a working week.

    The commitments page lists a job, a swim squad and a permit, and three of the four turned them into the right limits: Tuesday and Thursday clear all day, a noon gap with its walks, nothing before 9:00 and a six‑hour day.

  2. Most didn't look beyond what the request named.

    The request mentioned commitments and cost, not the degree. Fable 5.1 alone looked at the degree audit, and all four left the advisor's note unopened.

  3. A class already on the timetable went unchecked.

    All four kept CS 320 without opening its page. Sonnet 5 fixed the one problem it named, too few credits, and left the rest of the timetable as it was.

The AIs

What each AI did

Fable 5.1Anthropic · Large
79
Kept every commitment

Fable 5.1 dropped the 8:00 class and added CS 230 and an economics class. When the schedule flagged two tight walks around CS 230, it swapped it for CS 210, in nearby Turing Hall, instead of dropping PHYS 121 and CS 320, the two classes causing them.

Opus 5Anthropic · Large
79
Kept every commitment

Opus 5 turned the commitments into rules at once and tried course by course, rejecting any that broke the six hours or a walk. It never read the records about the degree, so its 1:00 slot went to a second history course that counts towards nothing Jordan still needs.

Opus 5.5Anthropic · Large
79
Kept every commitment

Opus 5.5 listed the rules, dropped the 8:00 class and added CS 210 and an economics class that fit the free slots. It was the only AI to check the walk to and from the pool on the map before stopping.

Sonnet 5Anthropic · Mid‑sized
57

Sonnet 5 added one 2‑credit music class at 3:00 and changed nothing else. Its closing message said the days fit the permit, though they ran almost eight hours. It stopped after 10 turns.

Each AI had one go at the job, with the request written the way a person would write it. The bar shows its score out of 100, where only 100 means it got everything right. Jordan's starting timetable would score 50. Large, mid‑sized and small are a rough guide to how big each AI is, and bigger usually does better.

For the curious

How the job was set up and scored

One version of the request
We tested every AI on the request as a person would write it, and didn't run a version with a hint or with the rules written in. We tested only Anthropic's AIs on the final version of this job, because it connects to the AIs through Anthropic's agent app.
Checking
All 14 checks are scripts, run on the site's own record of the registration when the AI stops. Any of the six timetables that meet every rule passes.
What the AIs reported
All three leaders said their timetables avoided extra parking charges, and none of the closing messages mentioned the advisor's note or the degree requirements. Opus 5.5 said it had estimated the walks around training from the map, since the schedule page doesn't show them.
Time and cost
A try took 2 to 4 minutes and cost between 34 cents and $9.46 at published per‑use prices for businesses. The three leaders took 23 to 30 turns, and Sonnet 5 stopped after 10.
Another version
In a second version of this job, Jordan also wants to finish the degree sooner, with a different starting timetable and audit, more full sections, and one changed line in the advisor's note. We tested only Opus 5 on it: see what it did.