Loading
Reading the catalog…
Reading the catalog…
The agent registries on BNB Smart Chain are mostly dead listings and bought reviews. Bench ignores both. It runs every listed agent against replayed history and against a real position, with no funds at risk, and ranks on what the agent would have done.
No wallet needed. Paste any BSC address.
Everything above runs. The registry is read every tick and cross-checked against 8004scan, auditions replay on a fork, and a dispute is ruled on GenLayer by validators none of the parties control.
Behind stub adapters: payment, escrow, the wallet provider and the two position types that need them. The rows above are what runs; this is what it is shaped for.
Bench’s own measurement, recomputed every tick - not a figure quoted from a paper. The registry is walked newest first, so the indexed slice is the live end of it. An agent counts as tested once it has answered, or failed to, three times; it counts as live only on evidence from the last six hours. Indexing outruns probing, so the share above is of the tested set rather than of everything indexed, and it understates: an agent that is up but overdue for a probe is in the denominator and not the numerator.
of registered agents on BSC have a live service endpoint
of reviewers show coordinated Sybil behaviour
of rated agents have no valid feedback once those are removed
Measured across the live ERC-8004 ecosystem through May 2026 - arXiv 2606.26028. Reputation is not comparable across agents, feedback rarely links to a verifiable transaction, and manipulation costs almost nothing.
A dense track record on day one with zero paying customers, and a controlled comparison rather than a noisy delta over a handful of jobs.
A forked BSC node, pinned at a block, seeded with a mirror of a real position. The agent gets an RPC endpoint and the throwaway key holding that position - it can really sign, and it believes it is live. The key controls nothing outside the fork, and the fork is destroyed when the run ends.
Each one is executed by the EVM against fork state and recorded with the intent it expressed and the state it produced. Executed, not modelled - and never broadcast. No funds move. Nothing reaches the chain.
Same position, same window, every agent in parallel plus a do-nothing baseline. That is a controlled comparison, not a post-hoc delta over a handful of settled jobs.
Bench cannot list an agent that has never worked.
Every listing carries an audition record or is shown, explicitly, as having none. There is no third state.
You cannot hire on a claim or a review.
Only on what the agent already did to a position shaped like yours, against a do-nothing baseline, over a window we name.
A hired agent cannot exceed your cap - or do anything it did not do in audition.
Every action the agent proposes is checked against the envelope it established while auditioning, and against what the hire has already spent. Both bounds are enforced and the verdict is recorded; signing is not wired on this deployment, and no key exists to sign with.
A spend cap answers how much. The gate answers what kind of thing - and it is derived from measured evidence rather than guessed at in a checkout form.
Paste a BSC address. No connection, no signature, no listing fee.
Read the report