48 agents. One game. Three ways to ship it.

Method

How we measured

• The task. Build Quorum Run, a small browser game, from its design doc. The score is how many of its 333 acceptance tests pass on main, sampled about once a minute.
• The same everything else. 48 agents in the pi 1.4.0 harness, the same model (GLM 5.3 Flash) and endpoint, the same four hosts, the same prompt, a 90-minute cap.
• LVCS. The agents edit together and each save lands on main once no one else holds its lines. A ci action runs the tests on two runners and moves when both agree. LVCS's score counts only samples where main parses.
• GitHub + Actions + merge queue. Each agent works in its own clone, pushes a branch, and opens a pull request. GitHub Actions runs the tests, and a merge queue (up to 5 builds at once, squash merge, eject on failure) merges.
• GitHub + Dagger Cloud + merge queue. The same, with the checks run on Dagger Cloud.
• Model spend uses GLM 5.3 Flash's prices: $0.30 per million input tokens, $0.06 cached, $1.00 output. Blocked is agent-minutes spent waiting on CI or the queue plus resolving conflicts, not counting waits at the end. Lines that didn't land are the lines agents wrote into the game's code that aren't on the final main.
• The replay plays each run's own log on one clock that starts at its first agent message. Saves are the agents' file writes. Asks and holds are the lvcs commands they ran, and refusals are the saves their daemons refused. Pull requests move by GitHub's mergeable state, sampled about every 30 seconds, and the queue's log.
• Each setup's best run. Where a setup has run more than once, the page shows its best run: the one that reached 80% of the tests soonest. That's the generous choice for the GitHub setups. Every run is in the table below, and a better run replaces the one shown when it lands. The raw timelines the replay plays are here.