Public evidence
Measure the organization, not one answer.
AgentOS publishes every declared benchmark attempt—including failures—and resolves reported results to sanitized evidence with exact subjects, environments and limitations.
3 of 5
declared quickstart attempts passed
16m 16s
fastest final accepted delivery
0
Captain follow-up turns in the final two passes
1 + 1
failed and incomplete attempts preserved
What the benchmark refuses to hide
It records human decisions, clarifications and repair; elapsed time, tools, retries and tokens; duplicated work; crash recovery; authority and chain of custody.
Missing telemetry stays unknown. Failed and incomplete attempts remain in the series. One proven Fleet is not presented as general production maturity.