Where web3 meets AI
I'm Adam Boudjemaa, a web3 x AI engineer based in France. I work where the two meet: token standards, tokenization and smart accounts on one side, production AI agents for regulated finance on the other. I'm a named author (1 of 5) of ERC-3643, the token standard used for $32B+ in tokenized assets, and I solo-shipped the initial 11-agent AI platform for a confidential US real-estate fund.
The full record: Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft. On the agent side, I also deployed about 46 agents for internal operations at Integra, where I was CTO (2025-26). Those are two separate systems, and I keep their numbers separate.
By Adam Boudjemaa · Published 2026-11-02 · Last updated 2026-11-02
// WHY THE INTERSECTION
Agents are starting to touch money and tokens
AI agents now call payment APIs, sign through smart accounts and ask to move tokenized assets. Each of those moments raises the same three questions: who authorized this agent, what is it allowed to spend, and who checks it before the transfer happens.
Tokens and agents answer those questions in different places. A permissioned token checks the holder. A smart account checks the signer. Neither checks the agent's mandate unless someone builds it. That gap is a focus of my work, and the rule I keep is simple: enforce permissions in the system, never on trust in the model.
// THE WEB3 SIDE
The standards I've written, and the code around them
I'm a named author on four Ethereum standards. Three are merged (ERC-3643, ERC-6960, ERC-7410). The fourth, ERC-8203, is a draft still in review in PR #1614, and I'm its second named author.
Before that, I was technical lead on Nexus at Biconomy, the modular smart-account stack built on ERC-4337 and ERC-7579. Biconomy's infrastructure processed 38M+ transactions and $441M in transfers. Smart accounts matter here because they are how an agent gets a scoped, revocable permission to sign.
Four Ethereum authorships, one still a draft
Status, role and spec link for ERC-3643, ERC-6960, ERC-7410 and the ERC-8203 draft.
ERC-3643 (T-REX), explained
The Final standard for permissioned security tokens. Cited by name by the SEC Chairman, which is a citation, not an endorsement.
ERC-6960, the Dual Layer Token
One contract for an asset and its classes or fractions. I'm the first author.
ERC-7410, allowance reduction by the spender
A small ERC-20 extension that lets a spender lower or revoke its own allowance.
ERC-8203 draft, PR #1614 (draft in review)
Agent Offchain Conditional Settlement. Not merged: a proposal, not a finished standard. Primary author Xianrui Qin.
How frozen tokens can still move in a batch
An ERC-7943 example that rejected 80 but accepted 40 + 40, and a proposed fix.
ERC-3643 on a Cosmos chain
Why a freeze can miss holders once a token crosses a VM boundary.
// THE AI SIDE
Production agents where a wrong answer is expensive
My AI work is agents and RAG inside regulated finance: answers that cite their sources and abstain when the evidence isn't there, agents held to least privilege, and evals that decide whether a release ships. The hub below is the long version.
AI in regulated finance, the hub
Evidence-grounded RAG, containment, permissions in the system and release-gating evals.
Case study: an 11-agent platform
The architecture and release checks behind the fund platform.
AI agents that move money
Why the permission check belongs in the payment system, not in the agent.
Testing an agent before launch
Go, wait or block: the eval suite that gates a release.
MCP without sessions
Where a stateless MCP server keeps what it used to keep per session.
// WHERE THEY MEET
Agent payments, agent identity, agent mandates
This is the actual intersection: an agent pays, a token checks, and someone later has to prove who approved what. These posts work through it one question at a time.
x402: linking each agent payment to its approval
x402 settles the payment. Delegation, spending policy and outcome records are still on you.
Agent payment standards: identity and spend limits
Who the agent is and what it may spend, including where the ERC-8203 draft (still in review) fits.
x402: how much value may come from AI agents?
What one x402 payment looks like, and why its record can't tell an agent from a script.
Know your agent: a mandate check on ERC-3643
Investor eligibility vs agent authority, and where the mandate check belongs.
// THE TOOLS
What I build for AI coding
I also ship open-source Claude Code plugins for writing, QA, design and security, plus selected standards implementations and Solidity tooling, including the ERC-8203 reference implementation (draft in review). Every one is on GitHub with its license and install line.
// FAQ
Short answers
What is a web3 x AI engineer?
An engineer who builds the systems where AI agents and blockchains meet: agents that request on-chain actions, tokens that check who may hold them, and the permissions in between. For me that means Ethereum standards work plus production AI agents in regulated finance.
Can an AI agent move tokenized assets?
Technically, yes: an agent can sign through a smart account with scoped, revocable permissions. Whether it may is a separate question. A permissioned token like ERC-3643 checks the recipient's identity and the transfer rules, not the agent's mandate, so that check has to be built somewhere else.
What is ERC-8203?
ERC-8203, Agent Offchain Conditional Settlement, is a draft in review (PR #1614) for AI agents that settle work conditionally off-chain. I'm a named author, second of three, behind primary author Xianrui Qin, and I wrote the reference implementation. It isn't merged, so treat it as a proposal, not a standard.
Where should I start reading?
Coming from AI? Start with the ERC-3643 page. Coming from web3? Start with the post on agents that move money. The blog's start-here page orders the rest by topic.