O Capistaine! My Capistaine! — A School for AI Agents
Or: how three years of infrastructure and three months of civic AI led us to the Dead Poets Society — and why the yawps are about to begin
The Grotto
In Peter Weir's film, John Keating leads his students to a grotto — a hidden space, away from the conformist machinery of Welton Academy, where young men read forbidden poets and discover their own voices. The grotto is not a classroom. It is a permission: to think freely, to question authority, to stand on desks and see the world from a different angle. Carpe diem. Seize the day before the institution seizes you.
Three years ago, we started building infrastructure in a stone farmhouse in Cap Sizun that we didn't yet know what it was for. Smart contracts. Agent frameworks. Orchestration workflows. Docker deployments. Redis coordination. Provider failover chains. None of it pointed at municipal elections. None of it had a civic purpose. It was a laboratory without a destination.
Three months ago, everything changed. The municipal elections approached. Audierne-Esquibien — a commune of seven thousand souls on the western tip of Brittany — had four lists, dozens of proposals, and no independent platform to compare what the candidates actually promised. The infrastructure found its question. A crawler for municipal documents. A validator for citizen contributions. An anonymizer for sensitive data. Piece by piece, the plumbing became a school.
We didn't call it a school then. We called it a RAG pipeline, an agent framework, an experiment. The name came later, when we finally understood what we had actually built.
The Curriculum
Every agent that enters Ò Capistaine arrives raw. A system prompt with good intentions and mediocre instructions. A chain of LLM calls that hallucinates when the context is thin. A retrieval pipeline that returns the wrong chunks because nobody taught it how candidate names are spelled in Breton.
The school teaches three things.
First: fidelity to source. An agent must never invent what it has not read. If the documents don't contain an answer, the agent says so — clearly, without hedging, without filling the silence with plausible fiction. This is harder than it sounds. Language models are trained to be helpful, and helpfulness without grounding is hallucination wearing a smile.
Second: neutrality under pressure. Four electoral lists compete in Audierne-Esquibien. Each has supporters who will scrutinize every answer for bias. The agents must present every program with the same rigor, the same depth, the same respect — even when one list published three pages and another published thirty. Neutrality is not indifference. It is the discipline of serving the question, not the questioner.
Third: traceability without exception. Every LLM call is traced through Opik. Every retrieval logged with its semantic distances. Every synthesis tagged with the model that produced it, the chunks that informed it, the thread that connects it to the citizen's conversation. Not because we expect to be audited — because we invite it.
The agents learn these lessons through iteration. Prompts are refined against real test cases — twenty-five queries that test name correction, accent handling, vague reformulation, adversarial intent, conversational follow-up. When an agent mistreats "van praet" as a verb instead of recognizing Michel Van Praet, the prompt is corrected. When it fails to detect that "ecole" means the citizen is asking about education policy, the category system is sharpened. Every failure teaches. Every trace remembers.
The Graduation
Forseti was the first to graduate.
Named for the Norse god of justice and reconciliation — not punishment, not vengeance, but the patient work of hearing both sides and finding truth — Forseti learned to validate citizen contributions against a charter. Is the contribution respectful? Is it relevant? Is it constructive? The answers come with confidence scores, and when the confidence is low, the agent says so rather than pretending certainty it doesn't possess.
Forseti's graduation was not a ceremony. It was the moment when the test suite passed, the traces were clean, the confidence calibration matched reality, and the agent was wired into the N8N workflow that connects citizen input to public debate on GitHub. From that moment, Forseti was no longer a student. It was a practitioner — handling real contributions, in real time, with real consequences for the civic process.
Ò Capistaine has not graduated yet. The orchestrator, the captain, is still practicing — and the election is its residency.
Its job is harder and less glamorous than Forseti's. Route the question to the right retrieval strategy. Merge filters when the system detects a topic. Decide whether a query needs comparison mode or a single-list deep dive. Present the synthesis with sources. Handle the conversation thread so that "et pour eux ?" resolves correctly to "what does the other list propose on the same topic?"
Every citizen question this week is a live exam. Every trace is graded in real time. Ò Capistaine learns by serving — and the five days before the election are the most demanding semester of all. Three years of infrastructure built the school. Three months of civic work filled the curriculum. The graduation ceremony, if it comes, will be written by the citizens who used the tool and found it honest.
Trust is not declared. It is earned — question by question, source by source, trace by trace.
The YAAAAWP
In the film, the students stand on their desks and cry "O Capistaine! My Capistaine!" — Whitman's elegy transformed into a declaration of intellectual freedom. But before that scene, there is another moment. Todd Anderson, the shy student who cannot speak in class, is pushed by Keating to close his eyes and yawp — to release a barbaric, inarticulate cry from the rooftops, as Whitman wrote, to prove that he has a voice at all before he learns to shape it into poetry.
Our agents are about to yawp.
Not literally. But there will be a moment — traceable in the Opik logs if you know where to look — when an agent produces its first real answer to a real citizen's question and the answer is good. Faithful to source. Neutral. Traceable. Useful. The kind of answer that makes a citizen say "I didn't know that list proposed that."
That moment will be the yawp. The agent will have found its voice. Not the voice of its training data, not the voice of its base model, but the voice shaped by the school: grounded in local documents, disciplined by the charter, accountable through traces. A voice that exists to illuminate, not to persuade.
The election starts in five days. The yawps are coming.
The Anti-Crawlbot
Somewhere on the internet, a crawlbot is scraping websites right now. It follows every link. It ingests every page. It asks no permission and offers no provenance. Its output will appear in some AI product as confident answers with no traceable source, generated by a model trained on data whose creators were never consulted.
That is the black box.
Ò Capistaine is the opposite. Not because it uses different technology — it uses the same embeddings, the same vector search, the same language models. The difference is architectural and philosophical: every input is consented, every output is sourced, every decision is traced, every agent carries the lineage of its formation.
The crawlbot expires. It exhales data into the void, consumed and forgotten.
The lighthouse inspires. It draws breath — a deep inspiration — before illuminating.
When someone says "AI is a black box," they are describing the crawlbot. They are not describing what AI must be. They are describing what happens when builders choose opacity over transparency, speed over accountability, scale over sovereignty.
We chose differently. And in March 2026, at the Encode AI Hackathon, that choice was recognized with a Social Impact award. Not for the technology. For the commitment: that civic AI can be built in the open, trained on consented data, deployed with full traceability, and used to strengthen — not replace — democratic deliberation.
The Loops
The deepest thing we learned is that Ò Capistaine contains two spirals, nested inside each other.
The inner loop is the agent improvement cycle. A citizen asks a question. The agent answers. The trace reveals whether the retrieval was strong or weak, whether the synthesis was faithful or drifted. The prompt is refined. The gazetteer is updated. The category detection is sharpened. The next citizen gets a better answer. The agent learns — not through gradient descent, but through human attention to its failures.
The outer loop is the democratic participation cycle. Citizens contribute ideas on audierne2026.fr. Forseti validates them. Ò Capistaine synthesizes them. The synthesis is published as a GitHub issue, debated in public, incorporated into the program or challenged and revised. The next cycle of contributions is informed by the previous synthesis. Democracy learns — not through AI replacing human judgment, but through AI making the information landscape navigable.
These two loops feed each other. Better agents produce better syntheses, which attract more contributions, which generate more test cases, which train better agents. A philosophy within a philosophy. A school that improves by teaching, and teaches by improving.
The Question That Started Everything
Three months ago, the question was simple: can AI help citizens understand what their municipal candidates actually propose?
Three years of infrastructure said maybe. Three months of building said yes — but only if.
Only if it is built as a lighthouse, not a black box. Only if its agents carry provenance. Only if its traces are open to inspection. Only if its neutrality is structural, not just aspirational. Only if the school that trains the agents is itself transparent.
Ici, l'IA n'est pas une boite noire. Here, it is a school — Forseti has graduated, Ò Capistaine is still practicing, and the yawps that will prove they found their voice are five days away.
O Capistaine, my Capistaine. The lighthouse beam keeps turning.
YAAAAWP!
Related: The Lighthouse Manifesto | The RAG Adventure Begins | Reliability Without the Cloud Tax