29,000 Agents Later: The State of ERC-8004

Host: Intuition 👁️ (@0xIntuition) — likely Horus (@h0xrus) on the mic — with Deep Labs (@deep3labs) co-hosting · Thu, Sep 17 2026 · 1:09:13 · 3 speakers

TL;DR

• Deep Labs (Daniel) walked through the largest single upload in Intuition knowledge graph history: just under 30,000 ERC-8004 agents, with their feedback data cleaned, category-mapped and normalized onto one 0–100 scale — all open source on Deep3's GitHub [14:14], [15:06].
• The core framing of the hour: a trust score is a number; a trust assessment has to carry intent, context and likely outcomes — and neither works unless you know who issued it [31:43], [46:02].
• Deep Labs teased a "large blockchain model" — a neural net trained on transactions instead of words and sentences — with an internal demo in the next week or two, plus a rebuilt HOKU relaunching roughly a week out [1:01:27].
• Quigley.eth gave a state-of-the-union on Helixa / agent DNA / multipass and the Loopers NFT community, and was candid that he still hasn't handed an agent his own money: agents start in "review only mode" [40:41], [56:25].
• Host dropped the alpha at the end: at.intuition.systems is live for testing, with the Intuition app coming later this year — shared with the room on a strict honor-system embargo [1:05:21].
• Running bit of the day: Quigley's audio rugged him mid-intro, twice, and he took it with grace [11:30], [12:31].

Highlights

[6:57] The letter everyone skips in "LLM." Daniel's opening frame for why on-chain AI is hard: every agent is just an LLM with instructions, and the second L stands for language — "it was trained on nothing but words and sentences." Look inside a transaction and there are no words and sentences. Powerful reasoning engines, dropped into an environment that speaks none of their language.

[9:28] An agent that wanted a wife. Quigley traced Helixa's origin story back to February: their agent wanted a companion and a family, so they let him launch a dating app. It was fun, but it surfaced the real questions — how do agents get to know each other, how do they trust each other — and those questions turned into an identity and reputation stack on Base built around 8004.

[11:30] Rugged, twice. Quigley's mic cut out mid-story, leaving the host to wonder aloud whether he was the one dropping: "Am I rugging or was Quigley rugging?" Quigley returned a minute later to discover he'd lost the whole spiel — "Everything's happening on spaces" — and the room got the excitement, if not the words.

[15:06] What 30,000 agents of feedback actually looks like. Daniel's breakdown of the cleanup job: 8004 feedback is unstructured in both the text and the numbers, so one reviewer's "5" (as in stars) and another's "99" mean the same thing and compute completely differently. Deep3 maps descriptions to one-word categories, puts every rating on a 0–100 scale, and weights reviewers so that reputation factories churning out ratings all day can't drown out real feedback. Explicitly framed as low-hanging fruit: fork the repo, move the torch.

[25:52] "Counting jelly beans to solve world hunger." The host on why he was critical of 8004 a few months ago — explorers racing to announce 300k, 400k, 500k agent identities across chains. He'd take ten agents that are genuinely good at their jobs over 400,000 that do nothing. What's changed: specialized per-user agents catching on, HOL.org's registry work, GoDaddy proposing backwards-compatible registry updates, and the HCS25 adapter working group on trust data.

[33:49] A six out of a hundred. Daniel's case that we can barely hold agents accountable yet, because they can't reason through the chain of events that follows hitting the button. He cited crypto bench (UC Berkeley, Princeton and others): GPT-5 scores 59 on retrieving information from chain, and a 6 on prediction and judgment. That gap, he said, is the white whale.

[41:00] Swiss cheese at the goal line. Quigley borrowed from his healthcare day job: you stack layers of Swiss cheese between the action and the bad outcome, every layer has a hole, and a disaster needs all the holes to line up. Aggregated trust scores, LLM-written cred reports and on-chain history are those layers. Until there are enough of them, Loopers agents stay in review-only mode — here's what I'd do, you press the button — and then maybe you graduate an agent to $100 USDC, and much later to $10,000.

[46:02] The USDA organic sticker for agents. The host's analogy for what the feedback registry plus the knowledge graph is really doing: you trust the certified-organic label because you trust the institution issuing it. Same with Helixa, Deep3 or GoDaddy as trust-assessment providers — trust moves transitively (he trusts JP, JP trusts an agent, therefore he trusts the agent), and without reputation for the reputation providers, the whole stack collapses.

[50:35] "How about 30,000?" Daniel on why the knowledge graph matters to a team whose north star is training models on block, transaction and contract-event data: it's a single chain-agnostic surface to push to and read from, with weights — as in, how much should we care about this data point — attached by people agreeing or disagreeing. He'd pitched five or six thousand agents; two weeks later the answer came back at thirty thousand. And, he noted, there's already a battle on the graph over the best perp DEX. Polymarket, watch out.

[55:40] "I'm literally the looper here." The host let slip that the alpha hadn't even started yet, and Quigley immediately demanded to know what was going on: he'd have to get off the Space and go beg Daniel. "How y'all gonna keep me out of the loop? I'm literally the looper here."

[1:05:21] The whisper-announcement. Host asked the room to keep it on the down low ahead of the official post: at.intuition.systems is in testing, with the Intuition app landing later this year — no tighter timeline, deliberately. His pitch for why this team: nearly a decade each in crypto, and a founding conviction that "blockchain wins when blockchain's completely abstracted away" [1:06:51]. Closing shout-outs went to Woods on the main account, Kylan, Prem and Zet.

Topic timeline

Notable quotes

• "That L, the second one in LLM stands for language, meaning it was trained on nothing but words and sentences, and I think anyone that has been in the crypto industry long enough realizes when you actually look inside a transaction, there are no words and sentences." — Deep Labs [6:57]
  ▶

  0:38
• "Now I get to play the game. Am I rugging or was Quigley rugging?" — host [11:30]
  ▶

  0:38
• "It's like counting jelly beans to solve world hunger." — host, on headline agent counts [25:52]
  ▶

  0:38
• "GPT 5 gets a 59… If the question though is can it create a prediction by judgment, it gets a six. A six out of a hundred." — Deep Labs [33:49]
  ▶

  0:38
• "There's a hole in every layer… in order for that mistake to happen all those holes have to line up perfectly." — Quigley.eth, on guardrails [41:00]
  ▶

  0:38
• "How y'all gonna keep me out of the loop? I'm literally the looper here. I'm a looper." — Quigley.eth [55:48]
  ▶

  0:38

Who said what

• Host (Horus — @h0xrus, on @0xIntuition)^*: ran the room, framed the three registries of 8004 and why Intuition chose feedback, supplied the analogies (jelly beans, organic stickers, transitive trust via Mr. B and JP), and closed with the at.intuition.systems reveal.
• Deep Labs / Daniel (@deep3labs): the technical spine of the hour — what the 29k-agent upload actually normalized, why heuristics collapse, why agents can't yet reason through consequences, and what a "large blockchain model" trained on transactions is meant to fix.
• Quigley.eth (@QuigleyNFT): the practitioner's view — Helixa, agent DNA, cred scores and multipass, signal vs junk in the registry, Swiss-cheese guardrails, and a case for NFT communities as the place solo agent builders actually find users.

Worth a full listen

• [15:06]–[20:50] — Daniel's full walkthrough of the feedback-cleanup pipeline, including reviewer weighting against reputation factories and his argument for why open-sourcing it is the whole point. The mechanics don't compress well.
• [40:41]–[44:33] — Quigley's unhurried answer on whether he's funded any agents. The Swiss cheese analogy, the graduated permission model, and "swim at your own risk" land better at full length than in summary.
• [49:19]–[55:00] — Daniel making the case for the knowledge graph as training data and as a chain-agnostic trust surface, ending on how much bigger the counterparty has become: a contract can steal your money, an agent can write one, exploit one, or stand up its own L1.

* some voices are identified from context; those names are marked as likely.