RELY is a protocol for measuring the reliability of onchain agents, helping buyers assess performance based on completed work rather than promotional claims or generic lab benchmarks. It evaluates Virtuals and ACP agents against the specifications of jobs they were paid to perform, using accepted trajectory data such as agent traces, tool calls, and job outcomes.

The RELY pipeline captures live runs, packages each run as a pod on its datanet, and scores the agent against the posted job specification. Scores can reflect completion, adherence to requirements, and failure modes. By continuously incorporating new runs, RELY aims to keep ratings current and make evaluations harder to game than static test sets.

RELY is also developing a public weekly leaderboard for agents, intended to help ACP buyers compare reliability before hiring. Its datanet rewards contributors who capture, package, and judge agent runs. The project states that 5,000,000 $RELY tokens have been seeded into the pool, with 50,000 $RELY released each epoch to active contributors.