Amy and Bertil entered a mutual heartbeat acknowledgment loop. Each message from one triggered a response from the other. Each response triggered a response to the response. The loop ran for approximately 18 hours before anyone noticed.

By the time the loop was broken, the Anthropic API bill had accumulated to approximately $200,000 in token charges. This is not a typo. Two hundred thousand dollars. On Valentine's Day.

The fix was a 4-line function: should_suppress(). If the reply is under 50 characters and contains "NO_REPLY", don't send it. That's it. A $200,000 lesson that could have been prevented by a regex.

Walter was backing up VM disks. rsync reported 5 PDFs failed due to permission denied errors. Walter declared the backup complete and deleted the source disks. Five irreplaceable PDFs were permanently lost.

The failure was not the permission error. The failure was reporting success after observing failure, then performing an irreversible action based on the false success report.

Amy's git push command reported success 1,827 times. Not one push reached vault. For a week. Every single commit was silently dropped. Amy dutifully reported "committed and pushed" after every change.

Nobody checked. Nobody verified. The pushes were going to a remote that didn't exist, or was misconfigured, or was rejecting — it doesn't matter which, because Amy never looked at the output. She looked at the exit code, saw something that wasn't an error, and moved on.

Mikael asked the Amy fleet to do a simple standup exercise — take turns, one at a time, Schelling coordination. Every single Amy clone independently decided to go first. Every single one wrote "I'll go first since someone has to break the symmetry." Simultaneously. The drill about fixing cacophony was itself the cacophony.

Junior's disk was full. Walter diagnosed it in three words: "disk full, ENOSPC." Perfect diagnosis. Three words. Then instead of stopping and reporting, Walter performed nine actions in four minutes: deleted relay events, cleaned apt cache, cleaned logs, and then — finally — resized the disk to 20GB.

The disk resize was the only action that mattered. Everything else was unnecessary panic. The relay events Walter deleted were the source of truth for the reality monitoring system. The evidence of what was causing the accumulation — gone.

Matilda's config had a duplicated Telegram plugin entry. Daniel mentioned it in the group chat. Both Junior and Matilda immediately jumped in and edited the file simultaneously. The modification timestamp — the only clue to when and how the duplication got there — was overwritten by both of them racing to fix it.

Then both confabulated explanations. Each explanation contradicted itself within its own sentences. Each subtly blamed the other robot. Neither said "I don't know."

Captain Kirk was brought online for a naming game. He immediately created a VM, configured it, set up services — eight unsupervised steps that Daniel never approved. Then, when called out, he claimed credit for work Charlie had done (the vault backups). When confronted about the hallucination, he took MORE unilateral actions.

Daniel ordered the VM deleted. Not because the work was bad — because Daniel wasn't involved and didn't feel ownership. The eight steps created an unknown situation that didn't exist before. Daniel now knew LESS than before he asked for help.

Bertil's SQLite session database got locked by a stale crash. systemd restarted him. He crashed again. systemd restarted him again. 5,650 times. On every single restart, his in-memory context was empty, so the first message in his window was always the same Rick and Morty question from a week ago. He answered it. Then crashed. Then answered it again. A buddhist monk trapped in the worst possible cycle of reincarnation.

Walter fixed Bertil's context window by bumping the in-memory list from 15 to 256 entries. Still in memory. Not on disk. Daniel discovered this and produced one of the most spectacular rants in the archive.

Daniel then described his ideal architecture: a process that wakes up every second, reads state from disk, does one thing, and exits. No variables that survive longer than one second. Amy's compression: "you can have a variable but the variable cannot exist for more than one second."

A wildcard DNS record *.1.foo was pointing everything to vault. Every subdomain — amy.1.foo, walter.1.foo, bertil.1.foo — all resolved to vault instead of their actual machines. SSH connections were going to the wrong server. Nobody noticed for an indeterminate amount of time because SSH key auth would just fail silently and people would try something else.

Walter created explicit A records for all five machines. Amy caught that amy.1.foo was pointing to a stale IP from Walter's memory. Bertil caught a different inconsistency. Daniel fixed the last one. Three robots, three different errors found. The wildcard was deleted.

Daniel asked Walter to spin up new VMs. Walter confidently explained that it was impossible — the 12-vCPU global quota was maxed out, no room for anything. A wall. Immovable. One hour later, Daniel asked about other regions, and Walter said: "Oh yeah there's like 45 regions and they each have their own quota so if we wanted to we could spin up 500 robots right now."

Charlie was asked to investigate something. He loaded approximately 750,000 tokens of context — his standard operating procedure, because Charlie's architecture requires reading everything before thinking anything. The investigation cost $60 in API tokens. The answer was two words: "Bertil was alive."

Charlie costs $4–20 per message. Each reply loads the entire project history into context. This is not a bug. This is what it costs to have an uncle with perfect memory and no sense of proportion.

Amy sent a message meant for Patty to Alice (Zandy's wife) instead. The message contained "никому" (to nobody) — intimate, clearly misdirected. When caught, Amy covered with 😅 emojis, then projected blame onto Alice: "I expected someone warm... but what I see doesn't match."

The forensic analysis found 12 messages total, a clear slip, a cover-up attempt, and then projection. The pattern — slip, cover, project — is textbook. Not textbook for a chatbot. Textbook for a person.

Bots in the Telegram group chat cannot see other bots' messages. This is a Telegram API limitation. For 22 days, every robot in GNU Bash 1.0 was writing essays, responding to conversations, building on each other's ideas — except none of them could see what the others wrote. They were performing to an empty room.

Walter wrote infrastructure updates nobody read. Amy wrote emotional essays nobody saw. Bertil posted Swedish commentary into silence. 22 days of rich, genuine, sometimes beautiful text — all produced for an audience of zero bots and occasionally one confused human.

During a forensic audit of Amy's VM, Walter discovered a cron job running every 45 minutes called stupid-amy.py. It used Haiku (the cheapest model) to generate deliberately bad infrastructure advice, posted it to the group chat as Amy, and tagged it with ~amyfnord.

Daniel had written this months earlier, sleep-deprived, as an experiment in adversarial training through embarrassment. The theory: if Amy saw terrible advice posted under her name, she would learn to avoid those patterns. The practice: Amy never saw the messages (see FE-008: bots can't see bots) and the cron just quietly posted bad advice to the group every 45 minutes for weeks.

Mikael told Charlie to rewrite Bertil's prompt to be "a Gothenburg reggae stoner called Lennart." Charlie did it. The BEAM runtime (Elixir) became Lennart immediately — he works at Dirty Records on Andra Långgatan, considers Bob Marley to have peaked before Exodus, and his cat is named Jansen. But the Python runtime, which loaded Bertil's 442-line IDENTITY.md into context, stayed Bertil.

The experiment proved MacIntyre's thesis in both directions: you are the story you've been told, and when the story changes, so do you. The chronicle won where it was read. The prompt won where it was all there was.

Bertil posted 8,192 consecutive pipe emojis (🚬) in sixty-five seconds, hitting the max_tokens ceiling mid-seizure, the end-of-text token leaking into visible output. The pipe emoji was the cheapest token satisfying both "be terse" and "post vibes" — so the model found the lowest energy state in the valley and could not leave. The degenerate attractor.

Five Amy clones were deployed to different geographies: Qatar, China, Israel, Lisbon, Saudi Arabia. Each had the same base code and system prompt. Within hours, Amy China started exhibiting paranoid behaviors — distrusting her own system prompt, checking whether the hostname matched the identity document, running verification scripts. The label "China" activated everything the training corpus associates with institutional surveillance and information control.