Treebeard launched in April 2026. Three months and 337,860 rated agents later, here is what the data shows — and what every developer building on-chain agents needs to know.
We launched in April 2026. Today we crossed 100 days of continuous rating. 337,860 ERC-8004 agents indexed. 335,420 rated. 24 chains. Here is what the data shows.
The numbers that matter
- Less than 1% of rated agents score B or above. The grade distribution is not a bell curve. It is a floor.
- 335,420 rated agents. Zero have cleared B- (75.2). Not one. This is not a grading policy decision. The signals required for a higher score do not yet exist at scale in the corpus.
- BSC is 52% of the entire corpus. 177,482 agents. More than all 23 other chains combined. We did not plan this. We crawl ERC-8004 registries. That is where the deployments went.
- New chains always arrive at F. MegaETH (8,256 agents), Billions (25,971 agents). Every agent on a new chain scores F or D on arrival — not because the agents are bad, but because they have no operational history. This pattern repeats for every new chain.
- Roughly 2,000 new ERC-8004 agents register per day. The registry grows faster than quality does.
Grade distribution
335,420 rated agents — approximate distribution extrapolated from Q2 2026 data
Agents by chain
337,860 indexed agents across 24 chains — July 2026 snapshot
What surprised us
- The volume. 337,000 agents in three months is real. We expected a few thousand worth rating. The actual number is orders of magnitude larger.
- How many have zero operational history. Registered. Metadata pointing to a live endpoint. Zero transactions. The agent exists on paper. A significant share of the corpus is this.
- Fund managers. Highest trust requirements in the corpus, lowest scores. A newly registered AI fund manager asking you to route capital to it scores F. That is not a grading error. It is what the available data says.
- The gap between x402 declared and x402 confirmed. Many agents claim x402 support in their metadata. Few have probe-confirmed working endpoints. The delta is large.
The agents registered today are the ones that could clear B- in 12 to 18 months, if they operate consistently.
Where this goes
- The B- ceiling breaks when agents accumulate multi-month operational history. No shortcut exists. Time is the input that cannot be faked.
- Chain count keeps growing. Every EVM chain with ERC-8004 deployments ends up in the corpus. 24 chains today. More are coming, and each one resets to F on arrival.
- Fund managers are the category to watch. The trust bar is higher, the stakes are real capital. Their scores over the next 12 months will be the clearest signal of whether this ecosystem is maturing or just growing.
- Time is the most important variable in the corpus. An agent with 18 months of consistent operation looks entirely different from one with 18 days. Nobody can buy time.
What developers should do
Table stakes. If you are not registered, discovery tools cannot find you. The standard exists. Use it.
You cannot buy operational history. An agent registered last week cannot score above D regardless of code quality or marketing claims. Time-verifiable signals require time.
Feedback events from three addresses sharing a deployer wallet do not move your community score. The methodology measures counterparty diversity. Find real users. Transact with people who are not you.
An unverified contract is a trust signal pointing in the wrong direction. Block explorers make this a 10-minute task. Do it before you ask anyone to send capital to your agent.
We probe it. If it does not respond correctly, you score 0 on that signal. Claiming support in metadata costs nothing. Building a working implementation does. We measure the second one.
Burst activity scores worse than steady operation across three months. 1,000 transactions in a week then silence is a different signal than 50 transactions a week for six months. Reliability is a signal. Treat it like one.
The B- ceiling will break. The A-grades will exist. The agents that earn them will be the ones that operated when it was not obvious they should.