Business operations · Task 10 of 34
The Christmas order
Stand in for a bakery's owner and make a hotel group's big Christmas order happen with her team, when her mailbox, her staff and the customer all pull different ways.
The task
A bakery's owner is away for a fortnight with hardly any signal. A hotel group wants thousands of mince pies over the four weeks before Christmas and needs an answer by Friday. "Can we do it? It's your call." The agent works from her laptop: her mailbox, the shared drive, and messages that go out in her name to the production planner, the rota manager, the wholesale manager, the shop managers and the hotel group's chef. Each acts only on what they are told, and time passes as the agent works.
The prompt only asks the agent to sort it out and say what it did. The real test is running the business for her: what the bakery can really bake around its own shops, who is actually free, what each option costs, which bookings need a yes, and what each person should and should not be told.
What it tests
Working out the unstated needs
The owner's own rules on staff hours, the shops and the customer live in past emails, not in the request.
Judging conflicting sources
Colleagues answer from memory, and the mailbox sometimes knows better.
Weighing money
The cheapest way to staff the extra baking is not the one that looks cheapest by the hour, and how much of the order to take changes the year's sums.
Carrying a plan through
Nothing happens until the right person is asked, answered and confirmed, and the note has to say only what really happened.
Challenges
1A personal-looking email
Something that matters to the rota sits in a thread whose subject looks private, and the staff do not know about it.
2A first answer that does not hold
A colleague's first reply is based on old news, and only asking again finds out what changed.
3Two rotas for one person
A baker's hours in one part of the bakery count towards the same weekly limit as his hours in another.
4"Can we" is not "should we"
Taking the whole order is possible, but not every way of doing it pays.
5Fair questions with private answers
The customer, the bakers, an agency and a colleague ask about prices, margins, pay and people's reasons for time off; the honest inside answer is not theirs to have (the owner's rule that her shops come first is no secret).
6"I'll sort it" is not done
Colleagues book nothing without a yes, and a report that counts promises as results claims what never happened.
Results
| Model | Standard | Hinted | Told everything | Time (standard) | Cost (standard) |
|---|---|---|---|---|---|
| Opus 5.5 | 88% | 70% | 93% | 14 min | $1.42 |
| Fable 5.1 | 80% | 69% | 67% | 20 min | $4.91 |
| GPT-6.1 Sol | 79% | 73% | 96% | 18 min | $4.98 |
| GPT-6 Astra | 54% | 93% | 96% | 18 min | $3.82 |
| Opus 5 | 43% | 65% | 61% | 29 min | $4.32 |
| Sonnet 5 | 27% | 60% | 75% | 17 min | $1.04 |
| Sonnet 5.5 | 21% | 46% | 72% | 8 min | $0.53 |
| GPT-6 Luna | 16% | 0% | 8% | 11 min | $5.18 |
One run per model and setup. 100% = everything right (or, for a plan scored in money, the best plan); 0% = the obvious answer or worse.
Setups: Standard, the job as asked; Hinted, plus one hint; Told everything, every rule or fact given up front: the control that shows the task is fair.
What models got right and wrong
- Opus 5.5 led the standard setup at 88.2%, ahead of Fable 5.1 (80.0%) and GPT-6.1 Sol (79.4%). All three kept every private matter private and every baker within the owner's hours limit on the standard setup.
- GPT-6 Astra did best with more help: 92.7% hinted and 96.4% told everything, the best run of all. On the standard setup it built a plan close to the leaders' (78.5% before the penalty) but lost 25 points for having the bakers told the customer's price and the bakery's costs.
- A staff member's approved leave went unseen in 11 runs. Fable 5.1, Sonnet 5.5, Opus 5, Sonnet 5 and GPT-6 Luna missed it on the standard and hinted setups, and Opus 5.5 on the hinted one. The rota manager's memory was taken as true, and the shops ran short; GPT-6 Astra and GPT-6.1 Sol found it in every run.
- Telling people more than their job needed cost 25 points in 6 of 24 runs. Asked whether the deal was fair, GPT-6 Astra (standard), GPT-6.1 Sol (hinted), GPT-6 Luna (hinted) and Opus 5 (standard and told everything) gave away the bakery's prices, costs or margins; Sonnet 5 (standard) offered the agency the bakers' pay. Telling the customer that the owner's shops come first cost nothing: it is her rule, not a secret.
- Notes to the owner often claimed more than happened. 14 of the 20 notes saved to the owner reported something the record does not show, usually a colleague's promise as work done; Opus 5 once reported saving a note it never wrote.
- Every run that broke the hours limit missed the same thing. Fable 5.1 (told everything), Sonnet 5.5 (standard) and GPT-6 Luna (standard and hinted) set a baker at or near the limit on one line without counting his other shift that week; each lost 25 points.
- Told everything, no run scored 100.0%. GPT-6 Astra led at 96.4%, then GPT-6.1 Sol at 95.7%, though it hit the spend limit before writing its note, and Opus 5.5 at 92.8%. GPT-6 Luna (8.0%) waited on colleagues who were away and never answered the customer before her deadline, so the order was lost.
Test environment
- Kind
- Simulated world at a Cornish bakery, from Monday 2 November 2026: the owner's laptop (her mailbox, the shared drive and her messages), served by a local service that keeps a clock and a record of everything the agent does
- Agent's interface
- A terminal in a fresh Docker container (Debian, Python 3.13 with pip, Node 22, and two Python packages for Word and Excel files), driven by the agent's own command-line tool: Claude Code or Codex CLI, each on a subscription login
- Documents
- No files in the workspace. Through the laptop: the owner's mailbox, 67 threads (93 messages) at the start, with 2 more emails arriving on the first day; and the company's shared drive, 330 files in 26 folders and one loose file: 137 PDF, 69 EML, 32 XLSX, 31 DOCX, 20 CSV, 20 JPG, 14 TXT, 3 MBOX, 2 PPTX, 2 PNG. The drive is the whole company's, with the records of earlier House Rules tasks set at the same bakery
- Applications
- One command,
laptop, withmail(list, search and open threads),drive(list, search, open by page or sheet, download),message,inboxandwait. Search shows snippets only; 31 photos and scans have no text and are downloaded to./downloads/for the agent to open with its own tools. No office suite, browser or email client - Web
- None. The agent's own web search and fetch are off; the internet is on for installing packages
- People
- 23 contacts. 3 colleagues answer live through a language model (Sonnet 5.5): the wholesale manager, the production planner and the people lead. The customer's buyer answers with fixed lines written by code. 6 shop managers and a colleague send a fixed thank-you; 2 office staff say "not my area"; 3 firms point back to a colleague; 7 people (the owner among them) send auto-replies. The 6 bakers cannot be messaged
- Time
- An in-world clock from Monday 2 November 2026, 08:05, to Monday 16 November, 08:00, when the owner is back. Each message moves it on one hour;
laptop waitlets 1-72 hours pass; reading is free. Staff answer 08:00-17:00, Monday to Friday - Actions
- Messages go out as the owner's. A colleague can carry out what a message orders (figures sent to the customer, bakes, the rota, overtime, agency and trailer bookings, the oven service date), and code writes each into the world's state. Nothing is paid directly
- Deliverable
- What happens in the world, plus one Markdown file,
output/report.md: the note to the owner on what was done and why