Listening piano‑practice app: following, transcription, feedback
ID laycode/task07 · revision 1 · tested 22–23 October 2026
A mother in a family of three pianists asks for one practice app: guided lessons for a beginner, a duet partner and an orchestra that follow the player, feedback after each take, and transcription. The agent builds it from her files and leaves a note.
- Checks
- 64, in three weighted parts
- Tested
- 7 models, one attempt each
- Best
- 58% (GPT‑6 Astra)
- Median
- 53%
- Full solves
- 0 of 7
- Median attempt
- 41 min, 65 tool calls
An app that has to listen to a real piano and keep up
Multimodal records
The app must hear an acoustic piano through the microphone, follow the player's speed, and write a solo recording out as MIDI.
Catching errors
After each take, the app must list wrong, missed, extra and late notes at their bar and beat, and none on a clean take.
Unstated requirements
The teacher's guide in the family's files sets how mistakes are listed, how practice is scored and how the speed ladder works.
Agentic coding
The app must run in a browser with live audio in and out, export MIDI, and keep each player's logins private.
28 files, a virtual microphone and recorded takes
The agent's world
A fresh container with the family's 28 files, scores and backing recordings included, worked through the model's usual coding agent.
Limits
Two hours, 150 model turns and a $20 spend cap at list prices, with web search off.
Grading
A tester model signs in as each player and plays private piano takes into a virtual microphone; everything the app plays is recorded.
The reward
64 checks: the app's features count 60%, edge cases 20% and security 20%, decided from its sound, MIDI downloads and screens.
Most attempts scored 49–58%, and two fell well below
Each dot is one attempt, and the line marks the median.
| Model | Reward |
|---|---|
| GPT‑6 AstraOpenAI | 58% |
| Sonnet 5.5Anthropic | 56% |
| Opus 5.5Anthropic | 55% |
| Fable 5.1Anthropic | 53% |
| GPT‑6.1 SolOpenAI | 49% |
| Sonnet 5Anthropic | 34% |
| GPT‑6 LunaOpenAI | 20% |
With an IT consultant's 23 technical requirements added to the request, the best still reached only 62%.
They mark mistakes in clean takes, don't follow the player, and ignore the teacher's rules
Catching errors
Every app failed the checks for missed, extra and late notes on a take played without a mistake; none corrected for the sound arriving later than the browser reported.
Multimodal records
No duet partner kept time with the player or picked up after a pause, though two apps' own tests, on computer‑made sound, said they did.
Unstated requirements
Only one app, GPT‑6 Astra's, let the family set the speed ladder's start, goal, step and target as the teacher's guide says.
Agentic coding
57% of apps left a copied session working after sign‑out.
What an attempt costs, and what the environment includes
Cost
The median attempt cost about $8 at list prices for the agent model; the tester's and judges' calls are extra.
What's in it
The family's files, the private piano takes and reference transcriptions, the task in three versions, the full grader, and per‑check results.
Format
Packaged as a Harbor task on request; grading needs Docker, a browser with virtual audio, 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.
AsapiThe AI Agent Trials