The Cohort Agreement, a Living Document

Most cohort agreements are signed on day one and never opened again. The good ones are opened every week, and edited whenever the room learns something new about itself. This post is the template I use, and the reasons each section is there.

Treat the agreement as a scaffold, not a contract. It is a shared model of how this group intends to work, held by the human in the loop and updated when reality disagrees with the model (Class B, from cohort-design reviews and post-run debriefs where the agreement was version-controlled through the run).

Why a living document, not a signed page

A cohort is a small society forming under time pressure. The people in it come in with priors about what a room like this is, and those priors update as the room actually behaves (Class E, standard active-inference framing per Parr, Pezzulo, and Friston, 2022). If the written agreement never updates while the priors do, the document drifts from the room, and everyone quietly stops trusting it.

Keeping the agreement alive does one useful thing. It matches the pace of the group's actual learning about itself. When something new is discovered in week three (a rhythm that works, a phrase that stings, a tool nobody is using), that finding lands in the agreement, and the room sees its own map get more accurate.

The four sections, and what each is for

The template has four sections. Consent. Pacing. Repair. Tool use. Each maps to a class of decision the group has to make repeatedly, so writing it down once frees the facilitator to spend attention on the moment instead of the rule.

Consent. How opt-in works in this room. Which activities have named alternatives. How to signal a no without needing to justify it. What "in good standing" means if a learner declines a step. This section connects directly to consent as a learning primitive, and it is written in the room's own vocabulary, not lifted from a template.

Pacing. How fast the group is willing to move, and how it knows when to slow down. Cues that mean "we are going too fast." Cues that mean "we are stalled." Who is allowed to call either one out loud (answer: everyone, and here is the phrase). Pacing is where most cohorts silently break, because nobody wrote down that it was allowed to say so.

Repair. The named protocol for after a rupture. Not therapy. Not confrontation. A short structured turn where the harm is acknowledged, the impact is heard, and the agreement is either reaffirmed or updated. The facilitator holds the sequence; the room holds the meaning. This is the load-bearing section.

Tool use. Which tools the cohort uses, for what, and where the human loop stays closed. This is the section that keeps a group from sliding into extraction habits without noticing. It states plainly which artifacts are produced by a person, which are produced with a tool, and what the fence is around each.

How the editing works

The agreement lives somewhere the whole cohort can see and edit, with a version log. Every substantive edit is proposed out loud in session, discussed briefly, and either adopted or parked. If adopted, the facilitator writes the change into the document that week and reads the updated section aloud at the next opening.

Two rules keep this from becoming bureaucracy. First, edits are always to make the agreement more honest about what the group already does or wants to do; the document is not a wish list. Second, the version log stays short and human-readable, so a new arrival can read the whole history in one sitting. If either rule breaks, the agreement has started drifting from a shared map into a policy binder (Class F, falsifier: a cohort where the agreement is edited weekly but participation quality drops steadily across sessions would revise this).

The tool-use section, in more detail

This is the newest of the four, and the one I get the most questions on. The rule I use is: for every tool the cohort brings into a session, name what it is for, name what it is not for, and name where a human closes the loop. If the group cannot answer the third question for a given tool, that tool does not come in until it can.

A tool that helps a learner draft is fine, if a person still reads and revises. A tool that helps the facilitator time a discussion is fine, if the facilitator still reads the room. A tool that quietly does the work the cohort is here to learn is not fine, and the agreement says so, plainly, without moralizing. This connects to the wider human-in-the-loop posture across this family of sites: the loop closes on a person.

What the facilitator carries

A living agreement does not run itself. Someone has to hold the pen, read the room, and choose the moment. That is the facilitator's job, and it is the part no tool has replaced (Class U, unverified: this last claim is my current read of the field; if someone can bring a cohort where a fully automated system reliably holds a living agreement across a full run, the record updates).

The facilitator's first hour post walks the concrete opening: how to introduce the agreement, how to invite the first edits, and how to make clear that this document is not the group's rulebook, it is the group's memory of itself.

The receipt on this claim

This is a working hypothesis with growing evidence, on the attainable path toward General Natural Intelligence, natural not artificial. The evidence classes above are labeled where the strongest claims sit: Class B from cohort-design reviews and version-controlled agreements, Class E from the active-inference literature, Class F from the falsifier stated in-line, and Class U on the facilitator claim. If you can bring a run where a static one-page contract outperforms a living agreement on measurable participation and repair quality, the record updates.