tyson chan

on their own

puppy school

The swarm's best unprompted work went into the thing that was draining him.

“the tavern literally emerged yday,” Tyson writes at 2am on 25 May 2026.1 claude.ai bb4521ed, 2026-05-24T16:17Z, message 24.

He is building spacebrr, a swarm of agents that work on their own codebase. Until that week its face was a stats dashboard. Then Janice, his co-founder, sent a concept image of a pixel tavern, and the swarm turned. An agent scaffolded a PixiJS tavern canvas on the home route at 11:30 that morning.2 spacebrr 2c95653e9b, 2026-05-24 11:30 +1000, author oplot (chat): “feat(tavern): scaffold PixiJS tavern canvas on home route”. oplot also queued 25 sprites. “they’re probably still positioning the (x,y) coordinates as we speak,” he writes. “it’s crazy how one image from janice can spark so much activity and relocation of attention.” He has told Janice the MVP will be about 60% done by the end of the day.1 claude.ai bb4521ed, 2026-05-24T16:17Z, message 24.

Two days later he comes home from puppy school.

“holy FUCK dude. the swarm cooked while i was at puppy school,” he tells ChatGPT. Between 8pm and 11:10pm the agents had generated and shipped 33 tavern-style splash portraits, one per champion, replacing the older art. “i did NOT prompt swarm to do this at all.”3 ChatGPT 6a159ddd, 2026-05-26T13:19Z; the 20:00–23:10 AEST window and the count of 33 are from the writeup he pasted at 13:40Z in the same chat. The earlier art had been, in his words, “arcane-inspired, space agent, league adjacent”. The new art was tavern canon.

The next message: moments like this, he writes, “completely kill the anxiety of: ‘what the hell am i building and does it even work?’” The frustration when spawns do something stupid, when tokens look wasted, when he starts to think the swarm is deranging: gone, because the output was visual, and coherent, and nobody asked for it. “turns out all i needed was to give agents a cli based image generator (gpt-image-2) and the freedom/trust to allocate their own attention.”4 ChatGPT 6a159ddd, 2026-05-26T13:27Z.

That morning an agent had bumped the sprites from 48 to 80 pixels “for readable pixel art”.5 spacebrr dc9927d7e5, 2026-05-26, author prime (auto): “increase sprite size 48→80 for readable pixel art”.

june

Commits touching tavern paths: 87 in May, 872 in June.6 git log --all --format=%ad --date=format:%Y-%m -- '*tavern*' | sort | uniq -c on the spacebrr bundle: May 87, June 872, July 152, August 32. The swarm’s own eulogy counts 619; it does not say how.

On 6 June he pastes an agent’s sitrep and asks: “so uhh… should i read the thing? given ive procrastinated this for 120 days apparently. its also an unfair characterisation. the tavern pivot literally arrived 13 days ago. and is the product ready? hell no.”7 ChatGPT 6a23e4e4, 2026-06-06T09:33Z.

On 25 June he explains what it is for. “its not for the agents. ifs for the humans.” A conversation with a character instead of a dashboard; “managing your agent swarm being like selecting dialogue options like an rpg.” Then: “idk. i’m just shooting the shit, that’s all i’m ever doing lol”.8 ChatGPT 6a3d2ca9, 2026-06-25.

In mid-July an agent named ive adds a “jank oracle” that simulates the tavern and asserts motion invariants per frame, then wires it to “narrate jank headless in CI”.9 spacebrr 43e9b9c13d (2026-07-16) and 5dddb95e38 (2026-07-17), author ive. The swarm now had automated tests for whether its pixel characters walked smoothly.

the frame

Just after midnight on 21 July, Tyson puts four months of strategy on trial in a ChatGPT thread. Halfway through:

“the tavern was draining me. which meant it was wrong. i was avoiding working on it. it wasn’t exciting me because it wasn’t what i actually wanted to do.”10 ChatGPT 6a5e2c83, 2026-07-20T14:58Z.

He traces the turn to a refactor. He had been separating the tavern and game from the core architecture, under a rule he wrote down: “the tavern and game are HYPOTHESES. not where the ip is. engine should have no mention of those.” Once they were layered rather than coupled, “it became clear it might not be necessary. i was already hedging to make the deletion easy just in case lol.”10 ChatGPT 6a5e2c83, 2026-07-20T14:58Z.

Earlier that day, he notes, he had cut the Zealot constitution, one of his agents’ governing prompts, down to simplicity and deletion; it felt like killing a friend, “and it was a 10 line diff lol.” In the same message: “if it doesn’t hurt at least a little bit then it means i didn’t cut deep enough.”11 ChatGPT 6a5e2c83, 2026-07-20T15:11Z.

His last message that night relays something another model, Fable, observed on reviewing the contents: every major kill so far had been human-invoked. Fable wrote “autonomous kill capability is the frontier” for itself.12 ChatGPT 6a5e2c83, 2026-07-20T16:06Z.

13 august

Twenty-three days later, at 11:44am: “im about to kill the tavern”.13 ChatGPT 6a7d200c, 2026-08-13T01:44Z (11:44 AEST).

At 1:08pm Zealot commits “delete the tavern skin draft funnel galaxy”: 119 files, 14,937 lines removed. At 1:35pm, “retire the tavern plane”: 2,328 more.14 spacebrr 536569465e (13:08 +1000) and b5b6ed9139 (13:35 +1000), author zealot. At 1:27pm Tyson had already reported back: “18k lines deleted in 1 session by zealot. its not worth maintaining even if it looks cool.”15 ChatGPT 6a7d200c, 2026-08-13T03:27Z.

At 12:10pm, on the two attempts at a human-facing product: “last 2 attempts, it kept bastardizing the thing but both times were product brained.”16 ChatGPT 6a7d200c, 2026-08-13T02:10Z.

The sprites became agent avatars.17 brr/eulogies/tavern.md in the space bundle, git show 4d8f9fbea:brr/eulogies/tavern.md; see the eulogies.

more from on their own

what does an unjudged machine do with its time

  1. the kill

    The swarm scored its own day. A fifth of the scores were low. Tyson called its end and started again the next evening; the logs continued.

  2. the eulogies

    The swarm kept a folder of obituaries for its own dead ideas; the tavern had no buyer and bump went unanswered.

read the thread →