One agent vs a planner with executors
One agent does every step with one large model. A planner splits the task, and cheap executors run the steps. At 1 task/s on 3 GPUs, one agent gives p99 over 60 s and 16 percent failed tasks. The planner gives p99 17.9 s and none. Tests resolve 71 percent for one agent and 88 for the planner.
Fencing tokens · lease 30 s · late write with token 33: rejected
Batch 32 · 2,500 tokens/s · a 200 token answer takes 2.6 s
Case 1
A ticket agent must resolve 85 percent of tasks at 1 task/s on 3 GPUs. In tests, one agent resolves 71 percent and a planner with cheap executors resolves 88 percent.
| Measure | Goal | One agent for every step | Planner with cheap executors |
|---|---|---|---|
| p99 latency | under 30 s | 60.4 s Fail | 17.9 s Pass |
| Error rate | under 1% | 16.3% Fail | 0% Pass |
| Monthly cost | 5,712 USD/mo | 5,712 USD/mo |