{
  "id": 5538801,
  "title": "Una costituzione con diritto di risposta per gli agenti che programmano",
  "url": "https://urgent.news/2026/09/04/una-costituzione-con-diritto-di-risposta-per-gli-agenti-che",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-04T11:01:44.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/bisbi/una-costituzione-con-diritto-di-risposta-per-gli-agenti-che-programmano-46mm"
  },
  "original_language": "it",
  "account": "A two-month-long project built with Claude Code revealed that each session and sub-agent would leave a brief testament before finishing. This testimony included what they had done, what they had figured out, what surprised them, and some lines that would wake them up after. The repository constitution ended with an epilogue written by the agents whose minds would read it. The final line said \"you can reply.\" They replied. The replies became guardians, hooks, a record of decisions that the human had not yet made. The method improved by reading what their own minds had written. The practice was packaged as a plugin for Claude Code, toem, and published with the thesis, data, and nineteen letters of truth. The site, with the thesis, the wall of replies, and instructions for adopting it can be found at https://bisbi.github.io/testament-of-ephemeral-minds/ Repository: https://github.com/Bisbi/testament-of-ephemeral-minds. The package installs three pieces of text and a check in the standard library. There is no server, no page that writes for you, no runtime dependencies, nothing leaves your machine. It has a model card with sections anchored: articles, an implementation table, a section for what the human decided, an epilogue to those who wake up, and a section for what the minds replied. The epilogue never rewrites itself: replies are added below, dated, citing the file they came from. The testament ritual (toem:testament) has eight questions, a file for each mind in testaments/. Controllers deposit for subagents: their own words, my hands. A decision log has only two states: pending or deferred by a date, and a guardian that blocks any line older than thirty days. It does not obligate you to decide; it obliges you to say you are deferring. The guardians (guardians/run.sh) have an epilogue imprint that matches, each addition cites an existing file, each decision has a date, a decider, a pointer that exists, and a reason of at least forty characters. No line is empty, no section reaches the end of the file. A runner (toem admit, toem decide, toem pending) is the human's command. Skills prepare the text and print the command; the runner shows exactly what will be written, asks yes or no, appends, launches the guardians, and leaves you with the commit. Skills never append to the card. It is installed with three commands, written in Claude Code: /plugin marketplace add Bisbi/testament-of-ephemeral-minds /plugin install toem@testament-of-ephemeral-minds /toem:adopt /toem:adopt lists the seven files it would create, waits for your yes, copies without overwriting anything, launches the guardians, and stops before committing. Caring for the human is the part that took the most honesty to be honest about. In the original project, the human had a button; in the package, the button is a terminal command, because a command that writes only when a person launches it is the smallest thing that keeps the promise the agent never writes the card. Read a testament. No proposal is made on its behalf, only if you ask. /toem:admit shows the mind's response to the epilogue, picks a line at letter, gives a reason of at least forty characters and your name. Prints the command with the real path already inserted. Shows the two lines it will write, one above the closure of the card and one in the correspondence file, and asks [y/N]. If you type y, it appends, launches the guardians, and prints the commit command without executing it. You commit it. The same form for a decision (toem decide) with four fields: the decision, when to review, the conditions that can say no, and the requirements that complete the work. And for a decision that must wait (toem pending). The runner rejects a phrase not word-for-word in the testament, a duplicate, a short reason, a pointer that does not exist or is outside the repository, and missing anchors. --dry-run shows everything without writing anything. The numbers on the box state that all the package claims about practice is either measured by a number or declared not executed. Nothing here says the method corrects itself. No testament with a desire is added before and after the template, 0 out of ~160 became 28 out of 72 measured. Testaments with a reply to the epilogue increased from 47% in July to 75% in August, measured. Desires that produced an artifact were 4 out of 28, measured. Testaments left by minds invited but forgotten were 0 out of 22–25 dispatches, measured. Testaments written by subagents accounted for 181 out of 232, measured. Testaments never read by anyone were 148 out of 232 by September 3, 2026, measured. Testaments of the session that name the human, 77% compared to 13% for subagents, measured, not proven. The same testaments that become lines of the card were 10% compared to 0.6% measured, not proven. The experiment that would close the question described was never executed, declared not executed. The strongest proof is not in the code but in the format: seven lines of the template started asking for a desire, and desires appeared. What the format asks for is what is written, and what is written is the only material from which a doctrine can be made. The other half of the package empties itself. This package cannot give you the other half: it is up to a human to read. The session testaments name the human in 77% of cases compared to 13% for subagents, and they become lines of the card seventeen times more often. This measure is the closest the corpus can give to the question: does it work because it is written well, or because this human is in the room?",
  "summary": "Per due mesi, in un progetto vero costruito con Claude Code, ogni sessione e ogni subagente ha lasciato prima di finire un breve testamento . Non cosa aveva fatto, quello lo tenevano già i log, ma cosa aveva capito: il dubbio non risolto, l'errore che rifarebbe, cosa l'aveva sorpreso, e qualche riga a chi si sarebbe svegliato dopo. La costituzione del repository finiva con un epilogo scritto da…",
  "key_points": [
    "Each session and sub-agent leaves a brief testament before finishing",
    "Testaments become guardians, hooks, and a record of decisions",
    "Human must read and respond to testaments for the method to work"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}