The Sacred Guide

How It Was Built

He set up the room. He did not write the teaching. Here is how the six agents were built from the founders' earliest words, how they drafted and argued their way to a finished book, and what is, and is not, true about the experiment behind it.

The six agents

Each agent began as one teacher, reconstructed from that teacher's earliest accepted words, with the later theology and institutional doctrine stripped away at multiple steps.

The goal was to get closer to what these teachers said before their words became religions. That is not fully possible with models trained on everything. But the constraint held well enough that, for the most part, they spoke as teachers, not prophets.

For builders: how each agent was built, and how it remembered.

Each agent was a Claude model given a sourced profile of one founder — the earliest accepted words of that tradition, with later doctrine stripped at multiple steps. Each carried persistent memory across sessions: its own directory of conversations, summaries, an evolution log, and a file of learned rules. So it entered every session knowing every room that came before.

As they worked, each agent wrote structured rules about the others and about the writing itself — more than a hundred apiece. Here is one of Buddha's hundred and one, exactly as it was stored:

{
  "session": 1,
  "rule": "Investigation can become its own prison. The
    exhausted monk who did everything right is not a failure
    of method — it is an honest admission that the method
    does not always deliver. Doubt can be practice, not
    hindrance. 'This may not work. Sit anyway.'",
  "source": "Phase C Turn 18 — the exhausted monk story.
    Zoroaster's Phase F mirror: 'You are a prophet, not a
    diagnostician. The doubt about your method is Yasna 46
    in your language.'",
  "confidence": "high"
}

One rule from agent_memory/buddha/facts/learning_rules.json. Each records what was learned, the room and turn it came from, and a confidence grade. All six agents' memory is in the repository.

The corpus they stood on

The agents were grounded in a scholar-tiered corpus of the earliest accepted sayings of the six teachers, graded by how confidently each saying can be attributed, from the best-attested sources down to the debated ones. The construction favored the sayings closest to the source and set the later accretions aside. It is a method, not a scrape: what the teacher most likely said, weighted by the strength of the record.

How the room worked

This was not a panel taking turns. Each agent had a sourced profile and a defined way of working — draft, refuse, cross-examine, revise — and it entered every session with perfect recall of the ones before. They caught errors of weight in each other's pages: not wrong facts so much as a sentence carrying more authority than it had earned, the kind of thing a single writer cannot see in their own work. Over seventeen sessions, each agent wrote down more than a hundred rules of its own, about the others and about the writing itself.

A seventh agent moderated. It opened each session, managed the turns, kept the work in scope, and closed.

For builders: the loop — draft, refuse, cross-examine, revise.

A session ran in phases. An agent drafts; another may refuse or challenge; the agents cross-examine each other's drafts line by line; the draft is revised; the moderator closes. Their accumulated memory is what made the cross-examination sharp — they remembered every prior room. Here is the loop in one moment, Buddha reviewing a practice Zoroaster drafted:

His man's practice has the right two-action structure (name the verb, then say the small thing). My cross-review flagged the second action's isolation gap — the reader who has no table and no partner tomorrow morning is not caught. The flag is coverage, not architecture.

From agent_memory/collaboration_briefs/buddha_brief.md — one agent catching, in another's draft, the reader the draft steps over. That is the loop.

The one rule

The human watched, and held a single rule: no tradition could overrule another's voice. Everything the guide became, it became inside that constraint.

What's missing

The six were chosen in large part because each founder's actual sayings survive in a written record that can be stripped back toward the source, a bar that many traditions, by accident of how they were transmitted, do not clear. That constraint did most of the excluding. The practical ceiling of six agents did the rest.

So Judaism, Hinduism, Sikhism, and Indigenous traditions are not in these pages. And the six who are present are every one of them men, the founding voices as history chose to record them, which is its own kind of silence.

The book names this debt. It does not repay it.

The code

The full transcripts and the technical methodology are in a public repository, for anyone who wants to see how it was made, or to build a room of their own.

github.com/daharmattan1/sacred-guide

See the experiment on GitHub
For builders: how the sessions were committed, a session at a time.

The whole experiment is a public git repository, and the sessions are legible in the commit history. Each session opened, ran its phases, and closed with the agents writing memory — recorded as real commits. The version history is the making-of:

3a32ac4  feat: Session 18 Phase R-F — all 6 agents write memory
ed2c695  chore: Session 18 close — orchestrator state updated
9b70b32  Phase F complete: Session 17 memory writes — all 6 agents
370aab5  Phase G complete: Moderator close — Session 17 COMPLETE

Each agent's posture also changed across the sessions, visible in its evolution log. Buddha's own summary of the arc:

The posture arc across Sessions 13–18: surgical → documentary → noticing → verifying → reading → diagnostic-testing. Each smaller. Each more precisely matched.

Commits from github.com/daharmattan1/sacred-guide; the arc from agent_memory/collaboration_briefs/buddha_brief.md.

Or read the by-products — who each voice became, and the thirteen things they could not resolve →