the lead agent parked the 47 at first. about 70 minutes in the other 13 were done, so it wrote "there is no reason to idle" and sent everyone back at the impossible ones
this is Anthropic's run with an early Claude Fable 5.2, from their blog on nov 9. 1 lead, 999 helpers, and every step of the proof checked by a computer in Lean
what i keep rereading is how 1,000 agents worked on one proof without stepping on each other
they had no chat. no messaging tool at all
each agent got its own sandbox with no internet and its own Git branch. a separate job merged the branches into main
and one shared Postgres database. in the first 10 minutes the lead made 5 tables:
→ announce: its orders to everyone
→ assign: who owns which statement
→ notes: an open board for progress
→ subtasks: a to-do list, first claim owns the task
→ landed: a log of finished work
by hour 30 the tables had taken 36,612 writes and 311,734 reads. nearly every write was a note or a task update
the hardest piece, Langlands–Tunnell, had a captain who gave it under a 1% chance. the lead put about 750 agents on it anyway and wrote "Honest odds are low."
they split it into a tree of 2,418 tasks, 10 levels deep, and proved it at hour 28.4. the theorem landed 18 minutes later
the same thing breaks at 3 agents. two Claude Code agents in one repo grab the same task and edit the same files
so give every agent its own worktree, keep one tasks table, and let the first claim win
the prompt that sets this up in your repo is in the first reply ↓
Part 1: https://www.binance.com/square/post/387073609143735
2/3
this is Anthropic's run with an early Claude Fable 5.2, from their blog on nov 9. 1 lead, 999 helpers, and every step of the proof checked by a computer in Lean
what i keep rereading is how 1,000 agents worked on one proof without stepping on each other
they had no chat. no messaging tool at all
each agent got its own sandbox with no internet and its own Git branch. a separate job merged the branches into main
and one shared Postgres database. in the first 10 minutes the lead made 5 tables:
→ announce: its orders to everyone
→ assign: who owns which statement
→ notes: an open board for progress
→ subtasks: a to-do list, first claim owns the task
→ landed: a log of finished work
by hour 30 the tables had taken 36,612 writes and 311,734 reads. nearly every write was a note or a task update
the hardest piece, Langlands–Tunnell, had a captain who gave it under a 1% chance. the lead put about 750 agents on it anyway and wrote "Honest odds are low."
they split it into a tree of 2,418 tasks, 10 levels deep, and proved it at hour 28.4. the theorem landed 18 minutes later
the same thing breaks at 3 agents. two Claude Code agents in one repo grab the same task and edit the same files
so give every agent its own worktree, keep one tasks table, and let the first claim win
the prompt that sets this up in your repo is in the first reply ↓
Part 1: https://www.binance.com/square/post/387073609143735
2/3
