Air traffic control for AI coding

Agents create output. Nool keeps their work coherent.

Nool against scheduled Git with 45 agents working in parallel: the question was not whether agents could write code, but whether their work stayed coherent when plans were wrong, noisy or incomplete. Four paired runs pooled across nool 7.0.3 and 7.2.0, not pre-registered, with the deltas the pricing page uses and nothing else.

• Acceptance rate 90.8% vs 80.4%: 218 of 240 proposed changes accepted against 193 of 240, paired runs only.
• 7.5 vs 18.8 conflicts per 100 agent attempts, about 60% fewer; 18 vs 45 merge conflicts across the four paired runs.
• 1.10 vs 1.24 agent attempts per accepted change, $0.134 vs $0.169 model cost, 0.083 vs 0.117 minutes wall-clock.
• Vendor-authored directional evidence from one workload: 7.0.3 and 7.2.0 runs pooled, not pre-registered.