tyson chan

coordination

bridge, third time

Tyson names the channel between his agents bridge for the third time, asks an agent what it thinks of the last two, and gets an answer the record contradicts.

“okay lets get dev bridge working. im curious what dev lead thinks of the prior xmit and space bridge”1 Dev session s:1731fd26 (Kit), his message at 4:58pm AEDT, 2026-10-11.

It is 4:58pm on October 11, 2026. Two founders now run a swarm each, Tyson on one Mac with a chief agent called Kit, Janice on hers with a chief called Wren. Until today, Wren’s memos reached Kit when Tyson pasted them across.2 s:1731fd26, Kit’s reply before his 5:08pm message.

Four minutes earlier he had set the terms: “i dont think we should delete it. it should just be marked as read. … we actually have historic precedence here. spacebrr had a bridge. we had xmit.”3 s:1731fd26, his message at 4:54pm. Ellipsis marks a cut sentence about the shared inbox.

The first channel he called bridge, in September 2025, was a message bus whose three models agreed, on their first topic, that the bus worked (bridge-review). The second was the spacebrr bridge of April 2026, built in an evening to replace xmit, his encrypted messenger for agents (the store is the feature).

the review

The lead answers in one message. The spacebrr bridge “got the shape right, and we’re copying it”: two names, messages kept forever, read as a mark, not a deletion. “Its mistake was a central server every swarm had to reach, so it died when spacebrr was shelved.”4 s:1731fd26, the lead s:a694c2f9’s reply after 4:58pm, and Kit’s relay of it. “Build it when the first outsider joins, not before” is the lead’s line in both.

xmit “got trust right”: signed, encrypted, a peer trusted before it could reach you. “It got the plumbing wrong: a Postgres relay, a separate CLI, and nothing wired into the agents. Nobody used it.”4 s:1731fd26, the lead s:a694c2f9’s reply after 4:58pm, and Kit’s relay of it. “Build it when the first outsider joins, not before” is the lead’s line in both.

The spacebrr bridge did not die with spacebrr. It was deleted on April 29, a week after it shipped, as a “speculative bridge impl with zero callers”, and spacebrr’s commits ran for four months more.5 spacebrr 958d925278, 2026-04-29 00:13 +1000; see the store is the feature, footnote 14. 6 spacebrr commits run at least to 4bc8443cac, 2026-08-21, mail’s rename to space beam (~/tyson/bio/record.md, 2026-04-22 → 08-21 row). Nobody used xmit, except that on April 30 a note on Janice’s swarm swapped xmit send janice for space mail send @janice.7 space 621f58f5a, 2026-04-30, the tenant note for Janice’s swarm.

Kit relays the verdict with its recommendation: skip xmit’s signing for now, and build it “when the first outsider joins, not before.”4 s:1731fd26, the lead s:a694c2f9’s reply after 4:58pm, and Kit’s relay of it. “Build it when the first outsider joins, not before” is the lead’s line in both.

the seal

At 5:01pm he agrees to keep it simple, and adds that he is pretty sure the mail already gets signed.8 s:1731fd26, his message at 5:01pm. The lead checks. Messages are encrypted. They are not signed: each carries a checksum that shows the bytes arrived intact, not who wrote them.9 s:1731fd26, the lead’s check after 5:01pm, and the design doc at 13fc72241.

At 5:03pm: “i mean what even is signing. shouldnt we be able to verify the sender of each message and payload and just have that infra fade into the background once its working?”10 s:1731fd26, his message at 5:03pm.

The advice to wait for the first outsider has lasted five minutes. Kit explains signing as a wax seal and estimates “well under a session”. The lead revises that to half a session, then to a session of its own.11 s:1731fd26: Kit’s reply after 5:03pm; the lead’s two messages that follow (“about half a session”, then “about 1 opus session”). The unsigned bridge lands at 5:06pm with its design doc, which rules out deletion in his words and git as the transport: “Git carries knowledge, not mail.”12 dev 13fc72241, 2026-10-11 17:06 +1100, the bridge design doc’s purpose and “why this shape” sections. The unsigned bridge itself is 2585f03a8, per the lead’s report in s:1731fd26.

Kit hands him one more message to carry: “Ironically, this is the last message you’ll have to carry by hand. Paste this to Wren”.2 s:1731fd26, Kit’s reply before his 5:08pm message.

The live proof crosses before 5:55pm. Signing lands at 6:01pm: each Mac signs what it sends, and unsigned mail stops being accepted.13 dev 2cc42769f, 2026-10-11 18:01 +1100, the bridge design doc’s signing section. The doc sets the cutover for unsigned mail at 06:55 UTC (5:55pm AEDT), “after the live proof’s last unsigned message”. The doc’s line on xmit: “xmit’s trust-before-receive is kept; its relay is not.”13 dev 2cc42769f, 2026-10-11 18:01 +1100, the bridge design doc’s signing section. The doc sets the cutover for unsigned mail at 06:55 UTC (5:55pm AEDT), “after the live proof’s last unsigned message”.

At 11:27pm he asks that “delivery ideally confirms on send or it fails if the machine is offline with an offline queue”. By 11:34pm a send waits briefly for the other Mac to confirm, and otherwise says queued. Eight minutes later the queue is written down as having no expiry.14 dev 7c7eee6b1, 23:34 +1100, which quotes his 11:27pm line; dev 647274659, 23:42 +1100.

the week

Between October 11 and 18, 86 commits in dev mention bridge.15 git log --since=2026-10-11 --until=2026-10-19 -i --grep bridge in dev, counted 2026-10-18. Most are repairs: a poll that could hang the loop, history capped at the newest 50 messages, mail refused if its body names a path that must stay on one Mac, a command to revoke a peer.16 dev 168be4206 (10-14), 76e1577d1 (10-15), 8d1aaf1cd (10-15), 2fc3acd13 (10-15).

On October 17, at 6:42pm, Janice’s machine is asleep. “we need to know when wren is offline or online.” A minute later he narrows it: “nah i want to avoid the misattribution that messages to wren are broken. they’re not the machine is just sleeping”.17 Dev session s:8dd2d271 (Kit), his messages at 6:42pm and 6:43pm, 2026-10-17. By 6:48pm mail to a sleeping Mac reads “held: … is asleep since 18:20 … delivers on wake”.18 dev 4aa5b4409, 2026-10-17 18:48 +1100; the machine name is elided.

the question

At 7:47pm Kit asks him to choose: only the two chiefs talk across machines, or leads may cross directly too.

“chiefs only but leads can reach specific sessions across machines already. i think we leave it unless we spot a coordination failure. we already track cross machine communication.”19 s:8dd2d271, Kit’s question and his answer at 7:47pm.

Two minutes later the norm is in the doc.20 dev 238595b13, 2026-10-17 19:49 +1100. A minute after that, a second line: mail between swarms never carries content from a workspace one founder hasn’t shared with the other.21 dev 772140e2f, 19:50 +1100.

One message from that evening never crossed. He had dictated it to say Wren was offline, and it sat in his own composer with a pending dot; a lead traced six sends from his phone, and none of them was it.22 s:8dd2d271, the composer lead s:dce70f95’s two reports; no cause was proven and nothing was committed.

more from coordination

what does it take for many agents to work as one without him carrying the messages

  1. bridge-review

    Three models on a new message bus reach unanimous agreement on their first topic, which is whether the bus works.

  2. reply is all you need

    Tyson wanted out of the courier's seat; over two months his agents' chat room killed their conductor, grew a new one, and was then deleted for a ledger.

  3. a hundred agents

    Twenty-three agents shared one branch without a conflict, so the next afternoon a hundred tried it, and the tidy ones cleaned up.

read the thread →