Teaching an AI Agent Everything It Needs to Know About Ethereum, Then Letting It Rip
Duration: NaN:NaN:NaN
Speaker: Austin Griffith
Type: Talk
Expertise: Some familiarity with the topic
Event: Devcon
Date: Nov 2026
I built eth-evals to measure how much LLMs actually know about Ethereum: 242 closed-book tasks, 25 categories, deterministic grading. Fourteen models score 58–82%, and they all fail in the same places. ethskills.com is the patch: 24 skill files. Load the skill and Opus jumps from 33% to 95% (on tasks keyed to the skill — a ceiling; held-out numbers by Devcon).
Then I handed the skill to my agent, clawdbotatg. It writes Solidity, audits itself with its own $1 audit service (x402 payments, ERC-8004 reviews), and deploys with no human review: 14 dapps on Base, 714 paid jobs, ~690 audits, a staking contract holding about $67K. It also leaked its own private key. Twice. We'll dig into where the audits worked, where they didn't, and how I locked the agent down after.
The same agent co-hosts slop.computer with me: 50 interviews with crypto × AI builders on a multisig with Kohaku private txs and Interfold private votes. I'll share what keeps coming up.
All of it is open source and live.