9 October 2026 · Heilbronn · case study
A learning day as a human-agent interface
One human intention, several agents, and a tutor that had to be steered back to the learner's level. This is what happened, marked by what was built, what was tested, and what was actually observed.
implemented tested observed in real use not yet observed
The day in commits · 11:00–21:00
44 commits, and a long gap
Each dot is a commit to the tutor repository. The marked span is when the tutor was mute because of a bug.
No commits between 14:22 and 20:40. Commit count verified in the repository.
Morning
The Wish at 11:26
Samuel wants a spoken tutor that explains concepts but leaves the thinking to him.
Applied AI student in Heilbronn, learning to program on his own. Second playthrough of a farming game driven by a Python-like drone, to understand concepts.
Morning wish: a spoken tutor that explains — but does not take the thinking away.
The repository was created at 11:26. repository created 11:26 · implemented The tutor itself is not yet present.
Morning
First Tutor, and a Human Voice
A working tutor exists within minutes, but the voice is a human decision.
11:33 the first working tutor is implemented and heard in real use. implemented observed in real use
11:36 the system voice is replaced; the local Piper voice “Thorsten” is observed in real use as a good voice. observed in real use
Midday
Midday: More Mechanics
The machine grows — screenshots, push-to-talk, cadence, didactic rules — and the coach talks a lot.
11:45–11:51 five sub-agents in parallel; four delivered with tests. implemented tests on four
Screenshots needed a small starter app (macOS permissions), observed from about 12:00. observed from ~12:00 12:38 push-to-talk with local speech recognition; questions recognized in real use at 12:55, 13:00, 14:02, 14:53 — one early question lost to a bug, then fixed. implemented observed in real use
Measured cadence: first ~62 s, later 32–42 s between statements. observed in real use
Afternoon
Afternoon: “Too Much”
The human corrects the coach: too much talk, reads code aloud, syntax annoys.
“Not on my level”: he often did not understand what Samuel intended; large variance; too many mechanisms — overengineered. (Human verdict.)
Reading code aloud and dictating a whole program were forbidden in the prompt and blocked in code. implemented tested Syntax hints capped at max 2 per hour. implemented tested
Afternoon into evening
Five Hours of Silence
The tutor goes mute through a bug, independent of whether the human is at the machine.
From 14:58 to 19:56 the tutor was mute through a hanging recording — independent of whether Samuel was at the machine. found afterwards in the logs
From about 15:00 to 18:00 there was no learning anyway: a break. An orphaned recording process held the microphone open almost 8 hours, writing into a deleted file; nothing was transmitted.
A 30-second limit plus a watchdog: implemented tested not yet observed again Model self-pausing ban — implemented.
Evening ·
The Human Thinks
The strongest moment is Samuel's own spoken reasoning about rounds and coordinates.
A nine-minute spoken thought note.
"I don't understand what a round is."
"In a round we run a 2x2 square and sow, then run it again and check."
"If I could remember and store which coordinates had dead pumpkins, and only check there the second time — I'd be more efficient."
Not a system claim — human reasoning. The plan was written down by Samuel and turned into code later.
Evening · –
Turning the Thought into Code
The plan becomes code — a borrowed 2x2 scaffold, then his own coordinate list, with open errors.
19:56–20:34 the plan became code. The 2x2 scaffold with a counter came from a separate chat assistant on request.
The coordinate list and targeted driving were his own implementation of the note.
2x2 scaffold · taken over from the chat assistant Coordinate list · written by Samuel
Evening · –
Radical Simplification
The review showed what actually helped — help on request, not the proactive tutor. So the coach is reduced: speak only on a question or on an execution error.
20:40–21:04 rebuild after the learner's verdict. The coach now speaks only when asked or when running a program throws an error; it reads the game's output and introduces itself once at start.
Implemented 20:47, automatically tested — never yet observed in real use (the game closed at 20:35).
Speak only on question or error · implemented + tested · not yet observed Concept card (24 concepts, 3 stages, editable) · implemented + tested · not yet observed Practice series (active minutes, from 25 min) · implemented + tested · not yet observed
Evening · into the coming days
The Loop, and Where the Rules Came From
The day ends as a repeating, deliberately undated learning loop built from the human's corrections.
One recurring daily appointment, 19:00–19:25, and a Notion path of 20 undated concepts: take the top open one.
- appointment
- top open concept
- own program
- touchstone
- streak
- next concept
The rules "no dictation" and "only on request" emerged only from Samuel's corrections — not in the morning.
44 commits by 21:04; 79 automated tests green.
Calendar + Notion loop · implemented "No dictation" / "only on request" rules · implemented + tested · origin: human corrections Commit and test counts · verified in the repository
End of day · still unresolved
The Open Problems
The unresolved problems remain part of the case study.
What HAI means here
Intention, execution, evidence, decision, updated state
The human steered through intentions and decisions; agents executed across system boundaries.
The goal was not automated productivity but infrastructure in which the learner's own ability can grow. The day's clearest finding came from the human: help on request worked, the proactive tutor did not.