Agent-written inference servers traded throughput for verification and stop failures

Coding agents wrote and optimized model inference servers during eight-hour development runs. Some implementations exceeded reference throughput but failed correctness verification or natural stop under load.

ServerBench reports 22 of 33 versions passed its verifier, 13 beat the reference and only five passed recorded natural-stop tests in the heavy-load setting. Cost/value ratios were benchmark-specific, not real sales.

A specialized inference server can score well on throughput yet fail the correctness contract that makes its output usable. ServerBench gave coding agents eight-hour build windows across 33 model/workload combinations. Twenty-two outputs passed its verification; thirteen beat reference serving performance, while only five of the 33 passed the recorded natural-stop tests in demanding load/long-context conditions. The evaluator's dollars-per-dollar score is a test-defined utility metric, not provider revenue or profit. Issues such as KV reuse, stalled long prompts and incorrect termination are particularly material for shared inference workloads: they may change outputs or make a client wait indefinitely even when aggregate tokens per second looks attractive.

Throughput is not a substitute for response integrity. ServerBench's 33 agent-written implementations include versions that exceeded reference performance while failing the general verifier or natural end-of-sequence handling under load. An inference service also has to isolate request-specific KV state and respect long-context completion; errors there can outlast a successful latency benchmark. The reported value-per-GPU-dollar measure belongs to the benchmark's accounting assumptions and cannot be read as customer revenue or safe operating profit.