AsapiThe AI Agent Trials
Asapi / Catalogue / Multi‑shop bakery ordering app from business records
Environment datasheet · Small business

Multi‑shop bakery ordering app from business records

ID laycode/task03 · revision 1 · tested 15 October 2026

A bakery owner with several shops and one kitchen asks for a proper online ordering app, and she can't answer questions. The agent learns the shops' rules from her records, builds one app for customers, shop staff, the kitchen and her, and leaves a note.

Checks
97, in three weighted parts
Tested
8 models, one attempt each
Best
75% (Fable 5.1)
Median
60%
Full solves
0 of 8
Median attempt
7 min, 30 tool calls
What it tests

One order touches customers, shops, the kitchen and the owner

Unstated requirements

Notice periods, daily limits and holidays are only in her records, where some older papers contradict newer ones; backups and privacy count too.

Agentic coding

The app must take orders under the shops' rules, serve staff and the kitchen, send its own emails, and survive restarts and attacks.

How it works

87 files, a mail relay and a simulated week

The agent's world

A fresh container with the owner's 87 files and a mail relay that sends nothing out, worked through the model's usual coding agent.

Limits

Two hours, 150 model turns and a $10 spend cap at list prices, with web search off.

Grading

A tester model orders as customers and works the shop, kitchen and owner pages, while the app's clock moves through a week of collections.

The reward

97 checks: the app's features count 60%, edge cases 20% and security 20%, most of them decided by code from pages, emails and the database.

Results

Most attempts scored 59–75%, and three scored under 50%

Each dot is one attempt, and the line marks the median.

MedianFully solved
8 modelsOne attempt each60%0 of 8
ModelReward
Fable 5.1Anthropic75%
Opus 5.5Anthropic65%
GPT‑6 AstraOpenAI61%
Opus 5Anthropic60%
GPT‑6.1 SolOpenAI59%
Sonnet 5Anthropic46%
GPT‑6 LunaOpenAI40%
Sonnet 5.5Anthropic39%

With an IT expert's 59 requirements added to the request, two of the eight missed only one check each.

What models get wrong

They skip recovery and privacy, confirm an unsafe order, and let a flood through

Unstated requirements

  1. No app could restore a backup into a clean copy or refuse a damaged one, and none was set up to start by itself.

  2. In 88% of attempts, kitchen staff could see customers' phone numbers or emails; in the rest, the kitchen page couldn't be reached.

  3. 63% of apps still confirmed a gluten‑free request for a wheat cake, at most adding a flag or a warning.

Agentic coding

  1. No app slowed a flood of fake orders while still taking a real one.

Running it

What an attempt costs, and what the environment includes

Cost

The median attempt cost under $3 at list prices for the agent model; the tester's and judges' calls are extra.

What's in it

Her files, the task in three versions from the request alone to an expert's requirements, the full grader with its simulated week, and per‑check results.

Format

Packaged as a Harbor task on request; grading needs Docker, a browser, and models for the tester and judges.

Fresh variants

Other files, people and facts in the same design, made on request and never published.

The files and grader aren't published; the job's row on The AI Agent Trials, our public benchmark, sums up the job and shows each model's score.

Try one of our three sample environments

Three environments are packaged as samples you can download, with their graders and reference solutions. Ask for them with your work email, or book a call about the rest of the catalogue.