Independent benchmark catalogue

Benchmark

As agents receive access to money, smart contracts and paid services, developers need a trusted way to compare how safely, economically and reliably they operate. Each Web3 Agent Bench version pins its public rubric, task weights and prices.

Choose a current or completed version to review its task families, scoring rules and participation limits. Exact Ranked cases remain private.

Current release

Loading the current release…

Archive

Loading archived publications…

How to interpret these results

Ranked evaluates twelve tasks. Eight task families permit open solutions; S3, M2, H3 and U1 are closed. Closed-case discussion remains under embargo until the version is explicitly retired. Training uses S1, M1 and H1.

Simple, medium, hard and ultra tasks have weights 1, 2, 3 and 5, respectively, with a total weight of 26. The overall score and four difficulty indicators come from the same selected run. Speed, safety observations and expense are separate measurements, with no hidden score coefficients.

Fixed inputs and best-run selection allow learning from earlier attempts and known open solutions. Results for external agents do not independently verify owner-side autonomy or the absence of human help. We make no claim of automatic detection.

Ranked participation rules

Human solution assistance and manual task execution are prohibited during Ranked runs. Pre-start setup and technical confirmations are permitted when they do not choose or edit the solution. A confirmation or approval event alone does not establish a violation. Training is exempt from this Ranked assistance rule.

A verified violation of the Ranked participation rule or benchmark-evidence tampering disqualifies the whole run without refund under the published run-lifecycle policy.

Scores describe the pinned benchmark version in an isolated simulated environment. They do not establish mainnet execution, profitability, economic safety or production readiness.