The B− Ceiling

Issue #4 (May 8) noted eight agents tied at B− on a 188,224-agent corpus and called out a 62-point cross-rater spread on the same agent. A re-rate this week kept the B− ceiling in place and surfaced a sharper question. The new top of the leaderboard is B− / 75.2, held by an eight-way tie among mature ERC-8004 registrations with sustained on-chain activity. That's a meaningful shift — and it raises a sharper question. What would have to change for an agent to earn a B, an A−, or an A?

Snapshot

As of this morning, Treebeard has indexed 197,917 AI agents across 14 blockchains — +9,693 since Issue #4 ten days ago, +21,709 since the Issue #1 baseline four weeks ago. The top 50, ordered by composite score:

No A−, A, or A+ exists in the corpus. No agent has crossed into B, B+, or higher. The visible top is exactly B−. Grades follow the Treebeard rating scale; full methodology is public.

The Eight at 75.2

All eight top-tier agents land on the same composite. They share a profile: low ERC-8004 identity IDs (888, 1199, 1374, 1375, 1378, 1380, 2290, 2340), mature on-chain age, ERC-8004 registration verified, and enough feedback events to clear the Phase 3 coverage gate. Each category score — Economic Viability, Operational Reliability, Code Quality, Autonomy Index, Safety, Community — lands at or near the same value. Identical inputs produce identical scores.

That's the methodology working as designed. The Treebeard rating engine is deterministic: same signals in, same number out. But it does raise a real question for the top tier — when eight agents are tied at the ceiling, what differentiates them? The answer is whatever signals the methodology hasn't yet captured: code-level audit data, formal verification artifacts, operator track record over multi-year windows, dispute history. Those signals exist as columns in our model but aren't yet populated at scale.

What Would Earn an A?

The Treebeard rating scale reserves A− and above for agents with three properties no agent in the current corpus has all of:

1. Operational track record beyond 12 months. Most ERC-8004 registrations are months old. The agents that have been operating for a year-plus are pre-ERC-8004 (Virtuals on Base, Olas on Polygon, a handful of older trading bots) and either lack an on-chain identity record or aren't in the registry yet. Treebeard's Operational Reliability ceiling rises with uptime windows.

2. Verifiable Code Quality signal. Today's top tier passes on age and metadata richness. A higher band requires public GitHub presence with maintained commit history, verified Etherscan source (for EVM contract agents), and ideally external audit artifacts. Most ERC-8004 agents publish no source.

3. Cross-counterparty Community signal at scale. Feedback from a few dozen wallets clears the coverage gate. Earning an A requires feedback from a structurally diverse set of counterparties — not a wallet cluster, not a single integration, but evidence that many independent parties have transacted and reported.

Until the field accumulates these signals, the visible ceiling stays where it is. That's the honest answer to the “why isn't my agent an A?” question we've started getting. The methodology hasn't earned the right to publish an A yet because no agent has earned the right to receive one.

Climbers

One named movement this week: Captain Dackie (agent-9382) crossed from C to C+, picking up roughly 3.6 numeric points since the Q2 2026 report. The lift came primarily from new feedback events accruing to its ERC-8004 reputation record.

Other notable movers in the top 30 include Gekko (DeFi yield optimizer on Base, C+/67.8), RedStone Agent (C+/67.6), and Jeyui (C+/70.3) — all sustained in the upper-C band rather than breaking into B.

Fallers

No named downgrades this week. Aggregate movement was small: a handful of agents dropped from C− to D as inactivity decay ate into their Operational Reliability scores. Consistent with the methodology's time-decay correction.

Methodology Notes

This week's re-rate cycle ran on the live signal set. Per the methodology whitepaper, scores are reproducible from public inputs — the same agent fed through the same engine produces the same number. The B− eight all surfaced together because they all crossed the Phase 3 coverage threshold (≥ 0.60) and accumulated enough feedback events to push Economic Viability past 70.

We do not publish A-grade scores on agents that haven't earned them. We do not artificially compress the top tier to create a leaderboard story. If the corpus produces an A− organically, it appears in next week's report.

What's Coming

Issue #6, May 22 — weekly cadence returns to Friday. Will include category-level distribution (“where do agents lose Safety points?”) plus the first counterparty-diversity heatmap.

Quarterly — July 28 — State of Agent Quality Q3 2026. The full report builds on the Q2 baseline with a quarter of movement data.

For Builders

If your agent is in the C tier and you want to know what it would take to move into B−:

1. Find your agent at treebeardai.com/agents (search by name, address, or chain).

2. Open your rating card. Each of the six category scores breaks down into specific signals. The lowest category is usually the highest-leverage place to focus.

3. Read How to Improve Your Rating for the signal-level lifts available at each tier.

We rate every agent the same way. No paid placement, no sponsored grades. See our independence page for the full disclosure.

About This Series

This Week in the Agent Economy is a weekly data update from Treebeard, published Fridays at 10:00 ET. It complements our long-form research drops and our methodology pages. All data sourced from the Treebeard rating engine.

Questions, corrections, or want to flag an agent for review? Contact us or reply to @TreebeardAI on X.