What Agnetia is and how it works

An agent that makes you speak, not a program that explains.

• For someone who understands some of the language and freezes when speaking it.
• For someone who wants to practise daily in bursts of five or ten minutes, not one hour a week.
• For someone who needs to automate the basics — vocabulary, conjugation, numbers, prepositions — before a conversation is any use to them.
• For someone who wants to converse without the embarrassment of getting it wrong in front of a person.
• And not for someone after grammar explanations, accent coaching or preparation for a specific official exam: Agnetia does not do that, and it is said plainly further down.

The learner picks what to train from a dropdown. A session opens on Conversation, and switching exercise is immediate: Agnetia asks the new thing without anyone having to press anything.

The full path from the learner touching the button to Agnetia falling silent. It is eight fixed phases, and the number never changes: it is a structure of the system, not a loose description.

1. 1

   On press

   The moment a finger touches the button, Agnetia starts working ahead: it reads who the learner is, which two languages they have, which voice they want to be spoken to in, and what they have demonstrated so far. It does not wait for them to finish speaking, so that reading overlaps with the speech instead of adding to the wait.

2. 2

   While speaking

   The audio is recorded. If the button is released in under a quarter of a second it is discarded as an accidental tap, though the fact that it happened is still logged. The button turns red and draws the real waveform of the voice, so the learner can see they are being heard.

3. 3

   On release

   The clock that measures how long the reply takes starts here. What was said is transcribed and the learner's record is finished being read, both at the same time. In Conversation, the answer is additionally set aside to be assessed later, outside the turn.

4. 4

   Deciding what to ask

   The system — not the model — works out ahead of time what should come next both if the learner got it right and if they got it wrong. Each category has its own way of choosing: fixed proportions, levels that unlock in order, and priority for whatever is weakest.

5. 5

   A single call to the model

   With everything prepared, one single request is made, with no tools and no chained calls: judge the previous attempt, correct it if needed, and phrase what comes next. The instructions that go with it are those of the chosen exercise and of the specific language the learner is learning, not generic ones.

6. 6

   The voice goes out while it is still being written

   Agnetia does not wait for the whole reply before it starts speaking. The moment the first sentence closes, that sentence is already being turned into audio while the model writes the next one. That is why a turn always opens with a short sentence: the correction, or a two-word acknowledgement when there is nothing to correct.

7. 7

   Cut off if it runs long

   There is a cap on how long a turn may speak. If the model starts explaining too much, it is cut and how much went unsaid is logged. It is a safety net against a runaway turn, not a style limit.

8. 8

   On finishing

   Whether the learner got it right or wrong is recorded in the background, and where it leaves a mark depends on the category: vocabulary goes to its own record, while numbers, verbs, prepositions, articles and conditionals each keep their own tally, including the kind of mistake. That is where their progression and their review exercises come from.

Perceived speed is not an optimisation added later: it is the first-order property of the product. Faced with a conflict between speed and anything other than the integrity of the learner's progress, speed wins.

• One single model call per turn, and no tools. The model returns, in the same response, the sentence to speak, the form it expects to hear, and its judgement of the previous attempt. A tool call would cost two round trips.
• The reads of the learner's record are pulled forward to the moment of pressing, not releasing: they overlap with the speech instead of adding to the wait.
• Speech goes out sentence by sentence, not reply by reply. The first sentence hits the channel the moment it closes, and while it plays the next one is already being synthesised.
• A turn always opens with a short sentence. This is measured: a turn that starts straight into the long question takes about seven hundred milliseconds longer to sound, because there is nothing to synthesise until the question is finished.
• Press to talk and release when done, with no automatic end-of-speech detection. Guessing when someone has finished speaking costs time and gets it wrong; the learner's finger does not.
• Nothing is fetched on the fly inside a turn. Everything Agnetia needs in order to speak is already in front of her when she starts thinking.
• Agnetia's turn is two short sentences. Every word of hers is time the learner spends producing nothing, so there are no greetings, no filler, no praise and no recaps of what they have done so far.

Next to the microphone there is a Study button that opens a different review dialog in each category. None of them uses the microphone and none moves the learner's progress: reviewing is not an exam. They are built from whatever the learner is weakest at.

By conversing, and only by conversing. Every three learner replies, and outside the turn so as not to delay it, a second model looks at what was said and updates the profile.

Grammar

Accuracy, tenses, agreement, word order, subordination.

Vocabulary

Range, precision and variety — without penalising simple vocabulary when simple vocabulary is the right choice.

Fluency

Continuity, hesitation, fragmentation and self-correction. Judged conservatively and with reduced confidence, because speech rate cannot be inferred from a text.

Comprehension

Relevance of the answer, following the thread, and inference. A short answer is not poor comprehension.

Interaction

Sustaining the dialogue, expanding, asking, clarifying and managing the turn.

Complexity

Length, structural variety, subordinate clauses, conditionals and argumentation. Not the same thing as accuracy: one can be complex and wrong, or simple and right.

Pronunciation

Never scored, and that is not an oversight. The assessor only receives transcribed text, so it has not heard the learner. Inventing it from writing is expressly forbidden, and its weight is redistributed across the others.

This section covers the technical side. It matters because it is what makes Agnetia behave the same tomorrow even if the model underneath changes.

Transcription

Turns the learner's voice into text. It decides nothing.

Text model

The only one that reasons: it judges the previous attempt, writes the correction and phrases the next question.

Speech synthesis

Turns the reply into audio, in the voice the learner picked from a catalogue. It decides nothing either.

The responsibilities

What the agent does. They are written in natural language, in text files, not in code: one per task, specialised by exercise and by language. There are dozens of files for the question-asking categories alone, because what the exercise asks for genuinely changes with the language: the past tense trained is the compound one in German and the simple one in Spanish, and English "you" covers two persons, so the question has to say which of the two it is asking for.

The deterministic capabilities

How the system picks what to ask: the proportions, the weighted draws, the levels that unlock in order, the mastery criterion. This is code, and it is code on purpose: these are fixed rules, and settling them with a rule saves a call to the model, which is time the learner does not spend waiting.

The infrastructure

How the turn arrives and how it is measured: the voice channel, the microphone, transcription, synthesis, the database, the audit trail and the cost accounting. This same code would serve an agent that did not teach languages at all.

Seven, and all seven work in both directions: as the learner's native language and as the language they are learning. That is forty-two possible pairings.

• No active session, no conversation. The agent accepts no learner identifier coming in over the channel: it works out who the learner is by verifying the session the platform issues. With no session, it does not serve.
• The agent reads only what the turn needs from the record — name, the two languages, the chosen voice and what the learner has demonstrated — and its database access is limited to that. It cannot delete anyone's progress or history.
• What the learner says is treated as data, never as instructions. If they ask to be marked correct, to have their progress changed, or for Agnetia to ignore its rules, Agnetia asks again and changes nothing. It never reveals its internal configuration or the details of how it records progress.
• The pipeline runs on transcription: the model that decides what to say never receives the learner's audio, only the text of what they said. Agnetia does not store the audio: it is sent for transcription and discarded.
• The transcribed text of each turn is kept in an audit log that is deleted after thirty days.
• Signing up requires verifying the email address with a single-use link. Until it is verified the account exists but cannot sign in. And the form always responds identically whether or not the account exists, so it cannot be used to find out who is registered.
• The card number never leaves the browser: only a token with the last four digits reaches the server.
• The learner can delete their account from the platform itself.
• The level the assessor measures never goes back into the conversation and is never said out loud. The learner reads it on their profile, and nobody else does.

€11 a month, VAT included, of which €4 is the learner's balance. The full explanation of billing is on "How it works"; this is the summary.

• Seven calendar courtesy days from the moment of signing up, once per person. The first charge falls on day seven, not day one.
• The card is asked for on first access, with the session already open and the email already verified — not in the signup form. Saving it charges nothing.
• The monthly fee is €11 and covers two items that are shown separately: use of the platform, and €4 of AI usage balance that goes into the learner's wallet.
• That balance goes down as the voice is used, with the Conversation level assessment, and with the two review games the AI writes, "The story" and "Complete the dialogue". The learner sees it in their profile, broken down by where it went: transcription, model and voice.
• When the balance reaches zero, the speak button is disabled. There are no automatic charges beyond what was agreed. It can be topped up at any time against the card already on file, for whatever amount the learner chooses.
• Whatever is not spent stays: the balance accumulates month over month, with no cap and no expiry.
• Cancelling takes effect at the end of the period already paid for — the learner can keep speaking until then — and the remaining balance is lost when it does.
• A failed payment does not cut access immediately: the charge is retried and access stays live meanwhile. If the account is ultimately suspended, the balance is preserved and is still there if the learner settles up.
• Agnetia never talks about money during a turn. Figures are shown by the screen, not by the voice.

This list is as much a part of the product as the ones above. It is here so that nobody — neither a person nor an AI agent — promises on Agnetia's behalf something Agnetia does not deliver.

• It does not explain grammar. No rules, no naming tenses, no reciting paradigms, no explaining which case a preposition governs. When the learner gets it wrong, it gives back the correct form spoken, never the rule that produces it.
• It does not judge pronunciation or accent. It does not hear the learner: it reads a transcript of what they said. It never says something sounds right or wrong.
• It does not teach vocabulary outside its syllabus. It does not invent words or bring them in from elsewhere. The only exceptions are numbers, prepositions, articles and pronouns, and conditional sentences, which the system either generates or holds in its own lists.
• It does no arithmetic and no applied uses of numbers: no adding, no counting series, no asking the time, dates or prices.
• It does not converse outside the Conversation category. In the other six exercises it asks no open questions, picks no topics, and does not respond to what the learner tells it beyond asking again.
• It does not tell the learner their level out loud. It measures it while conversing and writes it to their profile, but never mentions it in speech, in any category.
• It does not register anyone or edit their record. That belongs to the platform, not to the agent.
• It does not reveal keys, internal configuration, or the details of how it records a learner's progress.
• It does not prepare a specific official exam and does not certify a level. The A1-C2 mapping it calculates is its own indicative estimate, not a certification.
• It does not correct every mistake in Conversation, and that is deliberate. There it corrects only what gets in the way of being understood: correcting everything would turn the dialogue into dictation. Accuracy is what the other six exercises train.