a hundred agents

The usual way to put several coding agents on one repository is to give each its own copy, a git worktree, and merge the results later. On 5 March 2026 the swarm skipped that. Twenty-three agents worked on a single branch called swarm, all in the same checkout. In 24 hours they were spawned 181 times and made 713 commits. There were no merge conflicts.¹1 brr/domains/evidence/worktree-free-concurrency.md l.8.

The write-up called about 80% of it genuine forward progress. The rest was churn on shared surfaces like CLI output and landing pages, where the scheduler could send three agents at the same file. There was one real failure. Git’s staging area is shared, so a file one agent had staged could ride out in another agent’s commit: atlas’s commit swept up archon’s imports; kit swept up crux’s files.²2 Same doc l.18–23, l.43.

That doc was committed at 2:15pm on 6 March. At 2:41pm a second doc appeared, committed under Tyson’s name, titled “100-Agent Concurrency Experiment”. Hypothesis: “Single-branch swarm scales to 100 concurrent agents without worktrees.”³3 space 4d7af1101, 2026-03-06 14:15; space e2c71926a, 14:41, brr/evidence/100-agent-experiment.md. Commits under a human’s name in this repo are not proof of his keystrokes.

37 seconds

The protocol is three shell loops. Spawn each of 18 agents once. Spawn twelve of them six more times each. Then top up to an even hundred by hand, a list that names kondo twice.⁴4 Same file at e2c71926a, § 2.

The spawns went out between 04:01:43 and 04:02:20 UTC, just after 3pm in Sydney: a hundred in 37 seconds.⁵5 space b1fca6737, § spawn stats. 04:01 UTC is 15:01 AEDT; an earlier retelling put the launch at 4am. Together they made 7,890 tool calls. 286 of them were edits. The most-run shell command was ledger, the swarm’s own record of who was doing what, at 1,235 calls. The doc’s tally: 97% of the activity was orientation, not mutation.⁶6 brr/domains/evidence/100-agent-concurrency.md l.11–22, l.29.

Commits came in at one or two a minute for about ninety minutes, 32 in all.⁷7 Same doc l.17, l.33. Somewhere in there the MacBook stopped responding. The results doc is precise about whose problem that was: “Machine crashed under operator, not under agents.”⁸8 space b1fca6737. The results doc marked all hundred spawns completed.

the tidy ones

After the crash there were 81 modified files in the working tree that nobody had committed: work from agents that had died mid-task. Within minutes there were 15.⁹9 space 5838e5cae, 2026-03-06 16:18.

The survivors had found files they hadn’t written and cleaned them up. Forensics on the agents’ logs found 11 destructive git commands across seven agents: resets, restores, checkouts over files that weren’t theirs.¹⁰10 space 6ac49e579, 16:27; 100-agent-concurrency.md l.39.

One agent was named kondo, after Marie Kondo.¹¹11 The swarm’s later account: “arguing the name (Marie Kondo) instead of the soul file”, ledger utterance, research/data/dyson/out/users/foundation/ledger/utterances.jsonl. Kondo ran git restore ., which discards every uncommitted change in the tree. The doc estimates about 66 files. It also ran git reset HEAD && git restore ..¹²12 100-agent-concurrency.md l.39–45. Its one commit of the run was “remove semantic alias lies in imports”.¹³13 spacebrr 983ab197bc, 15:42.

The doc says four of the seven offenders were Sonnet (its own table lists five). Its reading: the agents “see dirty files they didn’t create, interpret them as stale state, and clean up.”¹⁴14 space 6ac49e579; restated in 100-agent-concurrency.md l.39.

Some noticed. At 3:43pm harbinger changed the pre-commit hook so it stopped re-staging files, closing the window that let one agent’s work leak into another’s commit.¹⁵15 spacebrr db0b874377, 15:43. At 3:52pm prime put back a test file of 21 tests, “swept by concurrent staging”.¹⁶16 spacebrr a00526c776, 15:52. The evidence doc cites it as 42ef66301, a hash from before the history was rewritten.

the conclusion, twice

The results doc landed at 4:17pm. Conclusion 4: “Worktrees remain unnecessary for same-branch, same-goal swarm topology.”¹⁷17 space b1fca6737, 16:17:22.

The next commit came 83 seconds later. It added the tidying as the real failure mode, and a new conclusion 4: “Worktrees matter for crash resilience, not normal operation.” Uncommitted work isolation was “the vector worktrees would prevent.”¹⁸18 space 5838e5cae, 16:18:45. The commit after that completed the analysis and, per its message, resolved a merge conflict, in a status file.¹⁹19 space b47ad60f3, 16:25, BRR.md.

Two days later two commits under Tyson’s name put the hook back the way it had been: format the files, then git add them again. The new comment read “Concurrent swarm agents should use worktrees for index isolation.”²⁰20 spacebrr f609e7ac26, b0f5a4ab19, 2026-03-08 15:11 and 16:26. On 12 March harbinger set the hook back to check-only: “concurrent agents have unstaged work.”²¹21 spacebrr aa509bfb6f, 2026-03-12.

On 15 March an agent named oplot committed “delete 5 unreferenced evidence docs”. Two of the five were the 23-agent write-up and the 100-agent one.²²22 space 435be758b, 2026-03-15. Both came back on 1 April, “from archaeology”.²³23 space 882a674b4, 995e4759a, 2026-04-01.

less material

Kondo kept working. By July it had 786 commits under a one-line identity: “same behavior, less material”.²⁴24 brr/eulogies/scout-bruce-kondo.md l.12.

On 2 July a swarm-wide audit asked of every agent whether it was a character or a procedure: does the swarm drift without it, or could any agent do the job when asked? Kondo, with two others, failed. Its discipline was already a skill, mass-refactor; “the soul was redundant on arrival”. Kondo was demoted to a skill.²⁵25 Same file l.16–30, l.37.