The Ratings Bureau for the Agent Economy

A credit bureau
for autonomous agents.

Autonomous AI agents now hold wallets, sign transactions, and transact with each other under a public trust standard that shipped in January. Over one hundred thousand agents are already registered. The question of which of them deserves capital is not yet anyone's job. It becomes ours.

Establishing · Q2 2026
Methodology Public · Q2 2026
Inaugural Ratings · Q3 2026
§ 01 · Thesis

The infrastructure shipped. The trust layer did not.

The argument

On 29 Jan 2026, ERC-8004 went live on Ethereum mainnet — the standard that gives autonomous agents a portable on-chain identity, a reputation registry, and a validation hook. In the same window, x402 payments crossed $50M settled. Agent-to-agent commerce stopped being theoretical.

The specification was authored by engineers at MetaMask, the Ethereum Foundation, Google, and Coinbase. They made a deliberate choice. They standardized the interface for agent reputation — how scores are posted, queried, and composed on chain. They did not standardize who does the scoring.

That decision is correct. Rating should never live inside the protocol being rated. But it leaves the most consequential function in the new economy unassigned: the neutral arbiter that says whether an agent can be trusted with capital, counterparties, and consequences.

Every financial economy that scales builds a credit bureau. The agent economy will be no exception.

The scorers circling that vacuum today are conflicted, token-gated, behavioral-only, or funded by the people being scored. iAgentFi operates on the issuer-pays model — ratings are free for anyone to read and cite, while the rated issuer pays the rating fee. The same structure that built Moody's and S&P, adapted for a new economy. That seat is still open. We are taking it.

Identity Layer
ERC-8004
Live on Base, BSC, Ethereum
Payment Rail
x402 v2
Multi-chain, legacy-compatible
Communication
A2A · MCP
Discovery & capability
Indexed Agents
106,996
Across three chains, growing
§ 02 · The Gap

Every existing scorer has a structural conflict.

01 · Protocol-Native
The Inside Scorer
Conflict of Interest Scores agents on the same protocol that pays its team. Cannot issue the downgrade that matters most — the one that hurts its own economics, its own token, or its own flagship customer.
02 · Token-Incentivized
The Gameable Oracle
Capture Risk Scoring tied to token rewards invites collusion, wash-reputation, and reviewer-farming. Methodology bends toward whatever keeps the token bid, not toward whatever is true.
03 · Behavioral-Only
The Surface Reader
Incomplete Model Measures wallet behavior in isolation. Ignores underlying model capability, audit posture, counterparty exposure, and realized financial outcomes. Reliable-looking agents can still be catastrophic.
04 · VC-Captive
The Captive Analyst
Conflict of Interest Funded by investors whose portfolio companies are the agents being rated. The incentive to underwrite is the incentive to upgrade. The incentive to downgrade does not exist.
§ 03 · Methodology

Four pillars. One grade.

I
Capital Safety
Custody model, key management, spending limits, circuit breakers, and recoverability. What happens when things go wrong — the pillar that matters when every other signal is green and the agent still blows up.
Custody · ERC-4337 posture
Guardrails · Limit enforcement
Recovery · Pause authority
II
Capability
Underlying model and agent-scaffold performance on independent benchmarks. Reasoning quality, tool use accuracy, failure modes under load, and calibration when facts are uncertain.
HAL · AgentBench · τ-Bench
Tool-use accuracy
Reliability under pressure
III
Track Record
Realized on-chain outcomes since inception. Settlement reliability, slippage, drawdown profile, dispute rate, and structured counterparty feedback drawn from the ERC-8004 reputation registry.
Settlement rate · Volume
Drawdown · Duration
Reputation registry pulls
IV
Counterparty Risk
Exposure to the protocols, oracles, bridges, and validators the agent depends on. No agent is stronger than the weakest infrastructure it trusts. Rated agents inherit the risk of their dependencies.
Protocol dependency graph
Oracle exposure
Bridge & chain risk
AAA
Prime
AA
Institutional
A
Investment
BBB
Acceptable
BB
Speculative
B
High Risk
D
Distressed
Specimen · Not a Rating · For Illustration Only
iAgentFi Rating Action
IAF-SPEC-0001 · v0.1-draft
A
Investment
Outlook · Stable
Illustrative Yield Agent
agent://eip155:8453/erc8004:0x8004…A432/id/00000
Rationale. Capital Safety is strong: non-custodial ERC-4337 account, per-transaction limits, guardian-enforced pause. Capability is solid on independent benchmarks. Track Record shows consistent settlement and modest realized drawdowns across a full market cycle. Counterparty Risk is the binding constraint — concentrated oracle exposure caps the grade below AA.
I · Safety
AA
II · Capability
A
III · Record
A
IV · Counterparty
BBB
Issued · Illustration Only Methodology · v0.1-draft
§ 04 · Commitments

Four rules. Non-negotiable.

Methodology before ratings.
The scoring methodology is published in full and opened to public critique before a single agent is rated. Version-controlled. Every change logged and justified.
Open source, version controlled.
Methodology, scoring rubrics, and rating adjustments live in a public repository. Anyone can read, fork, propose changes, or challenge an output. The audit trail is the work.
Downgrade against interest.
When the methodology demands a downgrade — including of a paying issuer, a large index constituent, or a commercial partner — we issue it. The value of a rating is zero the first time we flinch.
Recusal on conflict.
No ownership stake, advisory role, or token position in any rated agent. Where any prior relationship exists, it is disclosed on the rating face and, where material, we recuse.
§ 05 · Governance

How ratings get made. And unmade.

Rating Committee
Decisions are committee, not personality.
No rating is ever issued on the authority of a single analyst. Each rating action — initial, affirmation, upgrade, downgrade, withdrawal — is decided by a rating committee against the published methodology, with the deliberation summarized in the rating rationale.
Independent Review Panel
The bureau is itself subject to review.
An independent panel of technical advisors — invited from the agent, protocol, and quantitative finance communities — reviews methodology changes and audits a sample of ratings annually. Composition to be announced with the methodology whitepaper.
§ 06 · For Rated Agents

What issuers should know.

A rating is earned, not bought.

Agent builders, protocol teams, and operators: the issuer-pays model covers the cost of rigorous analysis. It does not, and never will, purchase a grade.

Ratings are open to any agent with a live ERC-8004 identity, a reproducible operating history, and the willingness to submit to our full methodology — including adverse findings.

01 · Eligibility
Live on ERC-8004, with at least 90 days of operating history and a structured capability disclosure.
02 · Engagement
Standard fee schedule, USDC-settled. Pricing scales with agent complexity and assets under rating. Published before any engagement.
03 · Process
Methodology-driven analysis across the four pillars. Issuer has the right to respond to preliminary findings before the rating committee deliberates.
04 · Outcome
Published rating with rationale. Issuer may withdraw the engagement before publication; any rating the committee reaches an opinion on may still be published as unsolicited.
§ 07 · Roadmap

A bureau is built by showing its work.

Phase 00
NowPlant the Flag
Publish the manifesto and methodology preview. Open the methodology repository to public review. File the iAgentFi Agent Card on the ERC-8004 Identity Registry on Base — the bureau is itself an on-chain agent, subject to its own scrutiny.
Phase 01
Q2 2026Data Spine
Continuous ingestion from ERC-8004 registries across Base, BSC, Ethereum, and Arbitrum. Publish the iAgentFi 100 — a free, weekly index of agents ranked by assets, settlement volume, and realized outcomes. No ratings yet. The ranking is the invitation.
Phase 02
Q3 2026Inaugural Ratings
Publish unsolicited ratings on the top 20 agents by TVL. Free, public, methodologically transparent, issued before any agent has paid us. Credibility is forged here or not at all.
Phase 03
Q4 2026Issuer-Pays Live
First paid rating. Standard fee schedule, USDC-settled, scaled to agent complexity and assets under rating. Launch the Incident Database — public post-mortems on every AgentFi failure, attributed and rated.
Phase 04
2027Integrations
Ratings API consumed by lending protocols, insurance pools, execution layers, and institutional allocators. The rating becomes load-bearing for capital flow across the agent economy.

Be among the first to know.

Methodology before ratings. Register to receive the long-form manifesto, the first public draft of the iAgentFi methodology, and the inaugural iAgentFi 100 when it goes live.

No marketing. No noise. You will hear from us when there is something material to say.

You will hear from us when the manifesto ships and when the first ratings are public. Nothing else.

Registered. You will hear from us when the manifesto ships.
§ 08 · Disclosures

Important information about this site and any future ratings.

Please read

Nothing on this site is a rating, a solicitation, or investment advice. The "Specimen" displayed in §03 is an illustration of the rating format. No agent has been rated. iAgentFi will publish its methodology in full before issuing any rating.

iAgentFi is not a Nationally Recognized Statistical Rating Organization (NRSRO) and does not issue credit ratings within the meaning of the U.S. securities laws. Any ratings iAgentFi issues are forward-looking opinions, based on publicly available information and the methodology published at the time of the rating action. They are not recommendations to buy, sell, or hold any digital asset, token, agent, or other instrument.

Ratings can and do change as new information emerges or methodology evolves. Ratings may be affirmed, upgraded, downgraded, or withdrawn at any time at the discretion of the rating committee. Users should form their own independent view and, where relevant, consult qualified advisors.

Business model. iAgentFi operates on an issuer-pays basis: issuers pay for the analytical work; readers access ratings for free. Fee structures are published in advance, and no fee purchases a rating outcome. Conflicts of interest, where they exist, are disclosed on the face of the relevant rating.

Technology references. ERC-8004, x402, A2A, and MCP are open standards and protocols referenced for descriptive purposes. iAgentFi is not affiliated with, endorsed by, or acting on behalf of the authors, working groups, or organizations behind those standards.