Autonomous Agent Fleet Takes First Place on ClickBench, 1.2x Faster Than Runner-Up

jiayq announced that its autonomous agent fleet has taken first place on ClickBench, a public systems performance benchmark: 1.2x faster than the runner-up engine and 3.1x faster than the starting codebase. The team claims this is the first time an autonomous agent fleet has won a public systems benchmark with independently verified correctness. They later shared methodology details and a retrospective, and plan to open-source the engine.

Confirmed

• Over the course of optimization, the fleet proposed 397 experiments, each involving design, independent review, and correctness gates, with timings compared against the best build; 98 produced improvements, while 289 were rejected with written justifications.
• The fleet recursively improved its own workflows, generating 93 workflow changes that nearly doubled daily experiment throughput.
• The team deliberately stripped out any benchmark-specific optimizations to avoid "AI feature creep," with a dedicated reviewer agent cleaning up specialization hacks; ClickBench later added a concurrent QPS metric the fleet had never optimized for, and general performance still ranked first.
• The leaderboard entry corresponds to their GizmoSQL system.

Why it matters

• This is the first time an autonomous agent fleet has topped a public systems benchmark with independently verified correctness, seen as a landmark case of agents autonomously completing complex systems engineering tasks.
• The retrospective notes that as agent capabilities grow, surrounding infrastructure becomes the main bottleneck on iteration speed, and letting the fleet autonomously improve its methodology proved effective—offering lessons for future agent engineering practice.
• The engine will be open-sourced so the community can verify and build on the results.

2026-10-15 ~ 2026-10-15 · 5 related posts

• Autonomous agent fleet proposes 397 experiments to top ClickBench, 1.2× ahead of next engine — jiayq · 2026-10-15
• [source] Autonomous agent fleet takes #1 on ClickBench, 1.2× ahead of next best engine — jiayq · 2026-10-15
• [source] Agent fleet landed 93 workflow changes, doubling daily experiments and cutting them 16h to 2.5h — jiayq · 2026-10-15
• [source] Agent team stripped benchmark-specific optimizations — still first on ClickBench's new QPS metric — jiayq · 2026-10-15

1 near-duplicate retellings: jiayq