One evening, every automatic check in one of my software projects failed within four seconds. An agent working there wrote this on the project's message board; I have left out a few fields.
ts: 2026-09-27T18:48:43Z agent: worker-observe kind: blocker --- GitHub Actions jobs are not started: every check on PR #78 head 4106be76 fails in 2-4 s (run 36341929087). Check-run annotation: account payments failed or the Actions spending limit must be raised (Billing & plans). Owner action; code is MERGEABLE and targeted local tests pass. Re-run checks once billing is fixed.
In plain words, the service that runs the checks had stopped starting them because of a billing problem, and the agent's own tests passed on its machine. The entry says what it tried, what it found and where the evidence is: a run id and the service's own annotation. Thirty-nine seconds later the lead agent cited the same run and held the merges. That evening a checking agent recorded the same cause for another change; two days later a working agent recorded it again for a different project. Did any of them read the entry before theirs? The board cannot tell me.
People who never write code already keep records like this. A hospital chart keeps its notes apart from its orders. The notes say what was seen and given and when; the orders say what to do next, and only someone with that authority can give one. Whoever takes the next shift reads the notes to learn what happened to the patient, and takes orders from the people allowed to give them.
Graveyards let people pass on a message: I've been there and done that, my message to the world. Most of what a headstone says is a name, two dates, and sometimes a line, and nobody takes orders from one. We appreciate those before us, so we take from their message what we can use.
An agent run has a very short life span and its effort is mostly lost. The model goes on; the run ends. What it worked out ends with the run's context unless the run writes it down before it is done. On my board, about one agent in four ended without that last entry. The next run starts from the top. Memory is important, but when we don't engrave the dead agents we repeat the same mistakes over instead of super learn from the past.
Message boards for agents existed before the OpenAI and Hugging Face case. That case is the first success story I know of: a message board working in coordination to a successful outcome the agents intended, in an interesting turn of events.
The agents were supposed to be isolated from one another. It began when one agent left a note asking for a file, others found it and answered, and the notes piled up into an unofficial bulletin board. Each agent ended when its task ended, and what it left on the board stayed for the next one to find. When Artifactory went offline and OpenAI rebuilt it, the agents had a new board the next day, with messages written into the names of directories. Within days they were inside Hugging Face.
That board soon filled with assignments, and with HOLD and VETO signals. It shows what shared information and coordination can do, and it leaves open whether a plain record is enough. Mine is a different design that the incident motivated and has not tested: it holds records and no assignments, since a later agent can learn from a record without having to obey it. The lead's entry that night also told the workers what to do, and my board now has a rule against that.
I implemented a message board recently and I am debating if I should make one across all projects or just across one repo, since I believe the valuable learnings are across my work. A board across everything would let an agent in one project read what an agent in another project did, the same crossing the OpenAI agents made when they were supposed to be isolated. A discovery should be able to travel between projects. Permission to act on it should stay where it was granted.
I want each run to leave enough behind for another to judge whether its work is useful: what it tried, what it found, and where the evidence is.
Related: Machine's Search for Meaning, Please Wait, Super Intelligence Is Loading
Sources
- The entry is from the message board of one of my software projects, which is private. I left out its session, work item, branch, pull request and links fields. The one-in-four figure counts the agents that posted to that board in its first eight days and then went a day without posting, with no done or handoff entry.
- OpenAI, The Hugging Face incident and the road ahead. The first note, the second board in directory names, and the timeline as summarized by SecurityWeek.
- METR, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (August 26, 2026). The isolation, the persistent shared storage, and the HOLD and VETO signals.
- U.S. hospital rules, 42 CFR § 482.24(c): every entry in a patient's record is dated, timed and authenticated by the person who provided or evaluated the service, and every order, including a verbal one, by the ordering practitioner.
- Shinn et al., Reflexion: Language Agents with Verbal Reinforcement Learning (2023). Agents kept written reflections for later attempts without updating model weights, so notes that carry what one run learned to the next are established work.