Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-20 · high
Ben Swerdlow built a Brood War interface for agents only and tracked dollars alongside wins.
Why now: Swerdlow's report carries the round-robin standings, current as of September 20.
Coding agents played a full round-robin of StarCraft: Brood War, with a price tag on every game, per engineer Ben Swerdlow.
The standings make that trade-off legible. The cheapest configuration in the top five costs a third of the second-place price, for a 16-percentage-point drop in wins.
Swerdlow built the game so it's playable only through agent interfaces. He then had every model and effort configuration play every other in a round-robin, on parallel Freestyle VMs.
Codex Astra at xhigh effort finished 18-0 for $10.54 a game, the only unbeaten run in the field. Codex Astra at medium effort went 16-2 for $15.11, more expensive despite losing more. Claude Fable finished 15-3 at $12.24. Codex 5.6 Sol at medium effort went 13-5 for $5.12.
Astra's medium-effort setting costs 43% more than its top setting and still loses two more games, the kind of pricing quirk a win-rate-only leaderboard hides. Sol at medium effort undercuts the rest of the top five on price, at $5.12 a game for 13 wins against five losses.
Brood War Bench's report lists every matchup but doesn't break down how the losses happened. Real-time strategy punishes slow decisions, and a 5-loss record doesn't say whether the cheap agent lost by inches or got overrun.
Each link below shares sources, entities, or timing with this story.
Boundary-Bench ran 12 agent harnesses through real firewall and filesystem locks, and costs climbed as much as 167 percent as those restrictions tightened.
Four Opus 4.6 agents coordinating through the system beat a solo agent on the newer Opus 4.8 model, 62.1% to 57.2%, per the July 30 paper.
Expert-authored code clears 82.2% of the same 53 tasks, with agent failures traced to skipped client communication and shallow requirement reading.
An ablation that skipped the router entirely tied the full system's score, per the paper.
A training-free fix called ChannelGuard held steady across three model backends, filter or no filter, blocking every tool-poisoning attempt.
Bun's Zig-to-Rust port spent 5.9 billion input tokens and 165,000 dollars, and passed the full existing test suite before merge.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.