Sherlock game shows small-agent swarm plus strong reasoner solves 96/100 murder cases

No_Yogurtcloset_7050 · reddit · 2026-10-11

The LMGame team built a procedurally generated Victorian London detective game (100 addresses, 10 suspects, one lead per address, must name culprit/motive/hideout within 12 case-hours) to test whether a swarm of small fast agents can solve murder cases.

Results (100 standard / 100 frame-up cases):

GPT6-Astra as Holmes + 12 Jev scout agents: 96 / 93
Rule-based Holmes + Jev swarm: 95 / 0
Lone GPT6-Astra doing everything itself: 57 / 59 (91 with 2x time, but over the limit)
100-leader Jev swarm voting: 38 / 42, median 2.2 case-hours (9s real time)

Key findings:

Swarms supply evidence coverage; the leader reasons and decides. Detection is divide-and-conquer.
Leaderless swarms agree too fast: confident wrong votes propagate before anyone verifies.
With honest evidence, a trivial clue-counting leader matches GPT6-Astra (95 vs 96) at a fraction of the cost.
Once witnesses lie, counting follows the bribed majority (0/100); only a reasoner that asks who provides evidence and why solves 93.

Open source, with other environments (CPU debugging, fish schooling, honeybee nest selection): github.com/lmgame-org/Gaming-JevSwarm

Original post →

Sherlock game shows small-agent swarm plus strong reasoner solves 96/100 murder cases

More from coding & agent