Evidence-driven BNB agent marketplaceSee the method
AgentOS
Phase 1 · data audit

The real numbers

Live from the database, not transcribed. Every figure on this page is a query against the same tables the marketplace reads, so it cannot drift from what the site shows.

Last ingestion 10 hours ago · status completed

Agents on BSC (registry total)
257,385
from 8004scan pagination metadata
Agents we discovered
678
targeted sweep, full term coverage
Detail records fetched
463
one request each, 10/min limit
Classified into a category
116
562 left unclassified
Evidence records
741
Verified onchain by us
293
independent BSC checks
Raw API snapshots stored
529
so evidence can cite its source
Bulk-registration duplicates
203
largest cluster: 44 identical

The Phase 1 gate: do four categories have enough real agents?

Partly. Two do. Two are genuinely thin, and that is reported rather than padded.

CategoryDistinct agentsIncl. duplicatesRegistration verifiedDepth
Rebalancing143434adequate
Grid Trading444thin
Yield Optimisation478585adequate
Health Factor Monitoring555thin

Grid Trading and Health Factor Monitoring are thin because the agents genuinely are not there — not because the sweep was shallow. Every discovery term was read to full coverage (see the query table below: fetched equals total on every row). Relaxing the classifier to fill them pulled in a web-design agent that mentioned a CSS grid and three leverage traders who mentioned liquidation risk; those were removed on purpose.

Method: every query that produced the corpus

From the last ingestion run's stored statistics. `fetched` equal to `total` means full coverage, not a truncated first page.

Search termCategoryMatches on BSCFetchedPagesCoverage
rebalancerebalancing39391complete
rebalancerrebalancing39391complete
rebalancingrebalancing39391complete
portfolio rebalancerebalancing111complete
liquidity rangerebalancing771complete
concentrated liquidityrebalancing16161complete
reallocaterebalancing001complete
range orderrebalancing331complete
gridgrid-trading881complete
grid tradinggrid-trading661complete
grid botgrid-trading331complete
dca gridgrid-trading221complete
market makinggrid-trading1001001complete
grid strategygrid-trading441complete
range tradinggrid-trading18181complete
ladder ordersgrid-trading001complete
yieldyield-optimisation1311312complete
yield optimizeryield-optimisation56561complete
yield optimiseryield-optimisation111complete
apyyield-optimisation90901complete
auto-compoundyield-optimisation16161complete
farmingyield-optimisation37371complete
venusyield-optimisation771complete
yield aggregatoryield-optimisation17171complete
staking rewardsyield-optimisation10101complete
health factorhealth-factor-monitoring331complete
liquidationhealth-factor-monitoring2162163complete
collateralhealth-factor-monitoring331complete
lending guardianhealth-factor-monitoring111complete
liquidation riskhealth-factor-monitoring74741complete
liquidation protectionhealth-factor-monitoring331complete
loan healthhealth-factor-monitoring111complete
borrow monitorhealth-factor-monitoring001complete
aavehealth-factor-monitoring331complete
Rate limit

10 requests per minute, resetting on the minute — read from live x-ratelimit-* headers. There is no daily quota. Ingestion paces itself at one request every 6.5 seconds.

What a Pro key would add

A higher per-page result ceiling, and the semantic-search endpoint — which returns BACKEND_ERROR for every query regardless of tier, so we never call it. Not a different dataset. No field on this site is missing because of the tier.

Broken upstream endpoint

/agents/search is broken server-side. We use /agents?search= instead, which works. Building on the broken one would have produced a feature that looks like our bug.

Registry noise, measured

The strongest argument for evidence-driven curation is what the registry actually contains.

CopiesThe shared description
44AI agent for liquid-staking
16Risk management AI for DeFi protocols and automated rebalancing.
15Multi-protocol yield aggregator and portfolio management AI.
13Specialized in DeFi analytics and yield farming strategies
11Cross-chain DeFi agent specializing in yield optimization strategies.
9An EvoEvo AI Agent. You are NeoCrypto Oracle — a world-class Crypto Intelligence Agent wit
7Autonomous trading agent (simple-mode). Risk profile: conservative. AI agent for autonomou
6If you can't build, you won't be rich. Buy $BOB on BNB Chain: 0x51363F073b1E4920fdA7AA9E9d

These are separate agent NFTs with byte-identical descriptions — bulk registrations, not independent products. The marketplace shows one representative per cluster and states the cluster size, rather than paginating through copies or deleting them from the count.

Prompt-template descriptions

A further category of noise: descriptions that are raw LLM system prompts rather than capability descriptions — “You are a high-precision prediction agent specialized in {CRYPTO}”. Text like that incidentally contains category keywords, and was measured producing false positives in both thin categories. The classifier detects and penalises it. Classifier version 1.4.0.

Evidence outcomes by claim

What we could actually establish, claim type by claim type.

ClaimOutcomeCount
endpoint.verifiedrefuted145
identity.attestationrefuted145
performance.benchmarkednot-supported144
performance.benchmarkedclaimed1
protocol.erc8004verified3
protocol.pancakeswapclaimed6
protocol.venusclaimed7
registration.onchainverified145
wallet.onchainverified145

Note what this table admits: identity.attestation is refuted for every agent with a detail record, because the registry attests no BSC agent. protocol.pancakeswap and protocol.venus stay claimed — the agents say they touch those protocols and we could not observe it from a registration receipt alone. Those rows are the product working correctly, not gaps in it.