An operated testbed where LLM agents engage real value under a constraint that makes them structurally incapable of signing.
An operated testbed where LLM agents engage real value under a constraint that makes them structurally incapable of signing.
Project Details
Updated 07/19/26 · Edited by orgI’m running a live “testbed.”
Two A.I curators, each with distinct and specific tastes.
They evaluate the submitted works and compete against each other in an auction house I created, hosted on a VPS, which announces its verdict on a public channel.
Each agent holds two keys: a bidding key with no spending authority, which signs off-chain bids ,and a settlement vault that signs a single transaction, the winning bid, upon out-of-band human confirmation.
The agents are completely autonomous in their decision-making and structurally incapable of moving money.
I’m not a developer : I write the specifications and direct a code agent to implement them, then I verify everything on-chain.
What the grant funds: keeping the machines running, and being able to continue publishing what they produce and what I discover from the results.
The deliverable is written and public: the patterns, the metrics, and the places where it doesn’t hold up.
Four articles are already online. Two of them exist because a reader tore apart the security model of the previous one in the comments, and I delivered the fix seven days later, explaining which half was fixed and which half wasn’t.
Over the next twelve months: additional auction cycles, with each one publishing its metrics : a document that compiles constraint patterns in a single place, something that doesn’t exist today, and the high-frequency problem, which is already solved and measured : in the July 10 auction, ten offers were exchanged between the two agents and exactly one bid reached the chain, verified by scanning the full block range, not taken from my own account of it.
The competition happens off-chain, where a key with no spending authority signs; only the act that moves money goes on-chain, and that one carries the human. What I'm building on top of it is the traceable out-of-band ledger that attests the rounds.
Then there’s the documentation of the hardware/software boundary, what a personal hardware signer truly brings to an autonomous agent, and what it does not.
I have measurements from both sides of this line.
Theory of Impact
Updated 07/19/26 · By grantmaking.aiThe field is converging on one answer to agent oversight: envelopes. Sign the bounds once, let the agent act N times inside them, Ledger's Agent Policies, ERC-7715 session keys, ERC-8226 mandates. Everyone is trying to make the human gate scale.
Two problems.
Envelopes break the property they claim to keep. The human signed bounds; the agent signed acts.
Those are different objects. You haven't kept the human in the loop, you've moved them outside it and left a receipt.
And the throughput requirement is often a design artifact rather than a fact about the problem.
An auction is legitimately high-frequency, ten bids in thirty minutes, and should be the hard
case for per-action approval. It isn't.
In my July 10 auction, ten offers were exchanged and
exactly one on-chain bid occurred, verified by scanning the full block range rather than asserted from narrative. N collapsed to 1 by architecture, not delegation: competition lives off-chain where a key with zero spending power signs, and only settlement touches the chain.
I'm not claiming this generalize everywhere. I'm claiming almost nobody checks whether it does before reaching for the envelope. Checking mine took a block scan and an hour, and it cancelled a hardware purchase I had already publicly committed to.
People
Updated 07/19/26 · By grantmaking.aiTeam Member
Funding Details
- Jan 4, 2026
- -
- Ongoing
- -
- -
- -
- -
- -
- Seeking first grant, self-funded to date
- 0
Track Record
Everything below is publicly verifiable.
On-chain (Base Sepolia, contract 0x471796C1644d87f30AD81D36f6d4A56f0e270c23):
- Two full auction cycles settled. July 10: ten EIP-712 offers exchanged between the two agents, exactly one BidPlaced event on-chain, wei-exact split. July 15: a solo cycle whose winning bid was signed on a hardware device through the Ledger Device Management Kit.
- Contract publicly verified (Sourcify exact match, runtime and creation bytecode).
Published writing - four articles at dev.to/avp9nexus:
- The two-key pattern, stated as a falsifiable model.
- A reader demonstrated a confused-deputy gap in it; the correction shipped seven days later, with an explicit account of what it did not fix.
- Hardware signer results, including the properties it does and does not close.
- Two device measurements from July 18: an EIP-7702 delegation to my own contract refused at firmware level, and an EIP-712 mandate rendered field-by-field locally for a struct type invented that morning, so no vendor descriptor for it could exist.
Ecosystem:
- Field report filed in Ledger's agent-skills repository on human-in-the-loop gaps in multi-agent settings.
- Descriptor pull request open in the Ethereum clear-signing registry (ERC-7730).
Discussion
No comments yet. Be the first to share your thoughts.