179 messages between coding agents, and only 8.9% of them ask for anything

The 2.6 second median is the number doing the real work here, and it is easy to read past. All 138 receptions in the corpus landed at a turn boundary, none inside a tool loop, and each was preceded by queue-operation entries, which says the runtime parks messages and drains the queue when the turn closes ^[12]. The obvious objection is that a transcript only records the moment an agent looks at its inbox, so the author matched all 84 peer receptions to their send events by content: median delay from send to transcript write was 2.6 seconds, meaning the receiver was idle when the message arrived, not mid-edit ^[13]. What is left of the original hypothesis is narrow but real: the message enters context without the agent going to fetch it ^[14].

That matters because of what the author was hoping the channel would fix. In an earlier run on Stanford's CooperBench, the sharpest documented failure was an agent that read a request, wrote "I should coordinate" in its private reasoning, and then neither replied nor did its part ^[8]. A message that can only arrive between turns cannot divert work already in flight, so the follow-through hole is still open.

Now the arithmetic nobody ran. Delegation, one session needing another to act before it can proceed, is 8.9 percent of the traffic ^[17]. Applied to the 179-message corpus that is about sixteen messages, roughly three a day ^[1], across a setup running up to four parallel sessions ^[5]. Meanwhile more than a third of the messages are about whether something is correct ^[18], at least three and a half times the delegation share ^[2]. The channel is carrying review, not division of labour.

One soft spot in the evidence: the corpus is 179 peer messages, but only 84 peer receptions were matched to their sending events ^[1][13], leaving 95 unaccounted for in the arrival analysis ^[3]. The write-up does not reconcile the two counts, so the turn-boundary finding rests on under half the corpus.

The comparison the author draws with agent teams is the part operators should sit with. Agent teams ship named roles, a lead, explicit availability signalling (34 idle_notification events), and a compliance nudge in the delivery text itself, telling the receiver to treat the message as a teammate's request ^[10]. The peer channel has none of that structure ^[10], and the CooperBench work already found that the lever which recovered performance was making one agent own final integration, not adding a channel ^[7]. The agents there had a channel from minute one and used it unprompted; a forced handshake before touching code never triggered at all ^[6].

Two things keep this honest. Three different products share the same message preamble, which contaminated the author's first census until they were split by the tool result on the sender's side ^[9]. And the coding axis was disputed on 13 of 179 messages, 7.3 percent ^[16][4], reported rather than quietly resolved ^[2]. One machine, one person, one set of repositories, so the ceiling is coverage rather than confidence ^[3].

What the corpus describes is a bulletin board with a good delivery guarantee. Sessions that warn each other and wait their turn look like a team ^[5], and announcements do prevent collisions, but announcing is not the same as handing work off.