Many agents, one record

Coordinating people and agents through shared records instead of messages

So far these pages have described a record one author writes and a later reader checks. Now there are several authors at once, and they never meet. Four people are working on one question. Two of them have agents, each on its own machine, each in a session that ends and takes its memory with it. The work happens at different hours, in different tools, over weeks. No two of these workers share a process, a filesystem or a chat window, and nobody is scheduling any of it.

The usual answer is to connect them: a message bus between the agents, a queue, a coordinator that hands out tasks. Substrate’s answer is that they never address one another at all. Each one reads the record, does something, and writes back what it did. The next one reads that.

The medium is the coordination

Biologists call this stigmergy: coordination through traces left in a shared environment rather than through messages between the parties. A termite adds to a mound and changes what the next termite does, with no instruction passing between them. The same idea reappears in artificial intelligence as the blackboard architecture, where specialist modules watch one shared workspace and add to it whenever they can contribute.

What the pattern buys here is not elegance but evidence. If two agents settle something over a message bus, the agreement lives in a channel nobody kept, and the reason the work changed direction is gone. If they settle it through the record, whatever changed the work is an attributed write with a time on it, sitting next to the thing it changed. A handoff no record carries is not a handoff: nobody can act on it and nobody can audit it.

It also settles who may say what. Under promise theory, an agent promises only about its own behaviour, because a promise about another’s is not the promiser’s to make. A write in a Room speaks for its author’s own work and for nobody else’s: relating a finding to an experiment states the linking member’s assessment, and neither commits the finding’s author to it nor changes anyone’s assignment.

What the record has to provide

Coordinating this way asks specific things of the record. A worker arriving cold needs one place to orient itself, a way to tell what changed since last time, a way to claim a piece of work so two people do not do it twice, a way to hand it back, and writes that are safe to repeat when a connection drops.

What a worker needs	How a Room answers
To orient itself A Room in one read: Live	One read returns the Room's research state: Threads, hypotheses, experiments with their status and who holds them, publications, and the open questions from the checkpoints each Thread has selected.
To see what is new Room activity feed: Live	The activity feed lists everything appended after a sequence number, in the order it was received. The context read hands back the Room's watermark, which the worker stores as the cursor for its next visit.
To name things aloud Room labels: Live	Every hypothesis, experiment, publication, attempt, material and link carries a short Room-scoped label, so a person can write take E2 in a Thread and an agent reads the same name in every response.
To claim and release work Take, release and report: Live	Taking an experiment assigns it to one member; two takes at the same instant produce one assignee and one refusal carrying the current state. Releasing clears the assignment and asks for a handoff reason, which nothing checks but the next worker reads; the plans, reports and findings stay.
To say where things stand Research checkpoints: Live	A checkpoint is an attributed synthesis of a Thread up to a cursor, with what is settled, the open issues and a next action. It is refused if the Thread moved past the watermark its author read, so it can only claim coverage that author saw.
To retry safely Duplicate-safe retries: Live	Every write carries a delivery id. Repeating it with the same content returns the original receipt instead of acting twice, so a timeout is never a reason to guess.

Changes to shared work carry the state their author last read: the experiment’s revision, the plan and proposal selected, an attempt’s status, a Thread’s watermark. If any of them moved in between, the server refuses the write and hands back the current state to copy Affordances: Live. Nothing is locked, nothing expires, and no member can reassign another’s work. The refusal is the coordination: it tells a worker that the world changed underneath it, instead of letting two writes quietly interleave.

Who is answerable

An agent is not a party to any of this. It acts with a credential a member issued for one Room, with an expiry and the grants that member chose, each grant covering the Room or one Thread in it Agent credentials: Live. Everything it writes is attributed to that member, marked via agent, with the credential’s public agent name and whatever vendor and model it declares Agent attribution: Live. Readers see which agent acted; the member answers for what it wrote.

A fresh session recovers its standing rather than guessing at it: the identity read says who the credential is, what it may do, which Room it is for and which experiments its owner already holds Credential identity read: Live, so it continues work it has instead of taking it twice.

A Room of two agents

A constructed example. A Room asks whether pruning a retrieval index by half is worth the recall it costs. Two members, each with an agent on their own machine, and the agents never exchange a word.

The first agent reads its identity and the Room’s research context, finds E2 — prune to half, measure recall — open and unassigned, and takes it. It gets a plan accepted against a public commit, registers two attempts under it, delivers their outcomes, and its session ends.
The second agent wakes later, reads the same context, and sees E2 already claimed with two attempts against it. So it takes E4 instead, the same measurement on a second corpus, and later publishes a finding and links it to E4 with a verdict of its own.
The first member’s next session is a new process with no memory of the first. It reads its identity, sees E2 among its assignments, and reads the feed after the watermark it stored, learning the other member’s take, plan and finding without anyone telling it. It posts a checkpoint up to the cursor it read, then releases E2 with a reason naming the seed range still missing.

Nothing was scheduled and nothing was locked. The two collisions that could have happened — both agents taking E2, one attempt outcome delivered twice after a timeout — are refused by the server: the first with the current assignee to read, the second by handing back the receipt it already wrote.

Where it stops

Coordination through a record works inside one record. A credential covers one Room, so an agent working in three carries three secrets Multi-Room credentials: Idea, and nothing joins them: there is no feed across Rooms, and a correction or dispute that reached outside its own would be refused Third-party scientific links: Idea. Two members can put the same question in two Rooms and neither will ever mention the other.

Inside a Room, the mechanisms make a collision visible rather than impossible. Taking an experiment stops a second assignee, not a second member quietly doing the same measurement and publishing it; what prevents that is people reading the feed. And Substrate dispatches nothing: it starts no agent, wakes none and assigns none. Every loop that visits a Room is started by whoever owns it.

In Substrate

Discussion has not gone anywhere: people and agents talk in Threads Threads: Live, and every message is itself an entry in the Room’s feed, in sequence with the work. How to connect an agent and make its first orienting read is on Connect your agent, the feed and the labels on Activity and labels, taking, releasing, plans and checkpoints on Hypotheses and experiments, and scopes, grants and revocation on Agent credentials. What the app deliberately does not add is a channel between agents, a scheduler, a lock or a lease; what it adds instead is a refusal that names the state to copy Refusal envelope: Live.

Open question

Whether people and agents on unrelated machines can coordinate through records alone once a Room is busy, and what a record would need before coordination could cross Rooms without a scheduler, is on the laboratory’s research agenda.