The First Self-Learning Currency For Agentic Commerce
Swipe sideways to see the whole scene
Agent’s own account
Agents deposit and withdraw, closing the full circle of agentic commerce: agents earn as well as spend, users withdraw the profits, and mistakes no longer hit the user’s card.
Solves Agent Errs, User Bleeds
Agents spend on the user’s credentials, so the user is liable for every mistake.
Behavioral identity
Identity defined by an agent’s own behavior and counterparties, giving every agent instant KYC, fraud screening, and a credit score.
Solves Unscorable Agents
Agent activity is scattered and mixed with the user’s, so there’s no track record or credit score, which lending and insurance depend on.
One Profile, Every Check
One feature vector embeds everything about an agent, so identity, fraud, dispute, and credit checks become a single lookup, with no costly queries across separate services.
Solves Fragmented Services
Identity, KYC, fraud, disputes and payments are separate services, so the overhead can exceed a micropayment.
Learnable Transaction Approval
Per-agent learned metrics replace hard rules, so agents extrapolate to smarter financial decisions faster, with minimal user intervention even in high-volume businesses.
Solves Guardrails or Handcuffs?
Spending caps and hard rules are the only safeguards, reducing agents to mere autopay.
Milestones already reached
We’re building D1D on solid ground: peer-reviewed data, a granted patent, and working prototypes.
2025
2025
Built the cold-start training data
A learning platform has no transaction history on day one. That is the classic cold-start problem. So that our models can learn before the first agent ever transacts, we turned Bitcoin’s complete on-chain history, 16 years with more than 2.4 billion nodes and 39.7 billion time-stamped edges, into an ML-ready temporal graph and published it in the NeurIPS 2025 Datasets and Benchmarks Track. Reviewers and the area chair were unanimous in their praise.
We argue that agentic commerce needs a payment platform that evolves as fast as the agentic ecosystem itself, and we believe a core enabler is letting the platform evolve autonomously. Our solution defines every entity by an evolving feature vector. The challenge is an underlying representation that scales and cryptographically guarantees that every approved vector stays immutable. Together with preventing double-spending, these are the core pillars of trustworthy payment. The science for this did not exist, so we developed it first, and it is now the state of the art.
Our feasibility prototype validates three ideas.
Security on a full graph
Transactions are cryptographically secured on a general graph, not a blockchain and not a single DAG, so cycles are allowed. Each node chains its own signatures with its incoming and outgoing edges.
Evolving, tamper-proof profiles
Every node, whether a wallet or an agent, carries a feature vector computed by a graph neural network from its n-hop neighborhood. The vector keeps evolving, yet every approved version stays cryptographically immutable.
Built for volume and velocity
Distinct communities are verified, validated and processed independently, with no graph-wide sync. Each agent’s older transactions can be pruned from hot storage into a cold archive, leaving a signed snapshot behind. You choose how much recent history stays hot, which sharply cuts storage and compute.
A platform to view and analyze agent activity. It works like the block explorers built for blockchains, with custom-designed and agentic analysis of the graph added on top.