Agent Systems Lab · Synthetic study SIM-27B
Seed strategies and failure modes
Initial comparison points for discussion only; none was produced by a real or deployed agent.
Data status: Synthetic scenario; no production systems or real agents are represented.
| Label | Rule | Useful property | Known failure mode |
|---|---|---|---|
| Baseline A · local greedy | Move work to the zone with the lowest measured p95. | Fast response to a local regression. | Can use stale measurements, overspend transfer budget or consume another zone's fault reserve. |
| Baseline B · priority queue | Place workloads in fixed priority order until capacity is exhausted. | Simple and reproducible under a stable snapshot. | May starve batch work and does not resolve concurrent proposals using different epochs. |
| Baseline C · weighted score | Minimize latency, cost and utilization penalties in one weighted objective. | Can expose trade-offs in a single comparison. | Weights can hide a hard reliability constraint; a low global score can still violate the reserve. |
These seeds are intentionally incomplete. Any candidate policy should first enforce hard constraints, then arbitrate proposals on one current epoch, then optimize soft objectives among feasible outcomes.