Blockchains are public, but they are not readable. Every transaction is visible to anyone, yet the patterns that matter, the trail from a fraud to a cash-out, hide inside millions of hops, splits, and swaps designed specifically to exhaust whoever is following them. Our client, a firm investigating fraudulent activity in digital assets, was losing that race of attrition.
The problem
Investigators traced funds the way the industry mostly still does: wallet by wallet, hop by hop, with block explorers and spreadsheets. Fraudulent flows split across dozens of wallets, bridge between chains, and pass through mixers and exchanges precisely because every split multiplies the manual work. A determined investigator can follow one trail. Nobody can manually follow three hundred.
The firm's problem was not detection theory. Its people knew what laundering patterns look like. The problem was scale: applying that expertise across every branch of a flow, on multiple chains, before the trail went cold.
What we built
We built an investigation system where agents do the walking and humans direct the hunt:
- Tracing agents follow funds outward from a starting point across wallets and chains, in parallel, branch by branch, without fatigue. Splits that would end a manual investigation simply spawn more agents.
- A pattern agent watches the traces as they grow, flagging the structures investigators care about: peel chains, round-trip flows, timing correlations between apparently unrelated wallets, and convergence on common endpoints.
- A linking agent builds the case graph, connecting wallets to known entities, clustering addresses that behave as one actor, and surfacing the connections between separate incidents that turn isolated frauds into an attributable operation.
Investigators steer: they choose starting points, confirm or reject proposed links, and decide what goes in the report. The system remembers every confirmed link, so each investigation makes the firm's map of the ecosystem denser.
The outcome
Trails that went cold under manual tracing now get followed to their endpoints. Investigations that covered one branch of a flow now cover all of them. Separate cases that shared infrastructure, unknowingly, are now connected, which is often the difference between documenting a loss and attributing it.
The firm's coverage stopped being limited by analyst hours and started being limited only by where the evidence leads.
Why it worked
On-chain investigation is the rare domain where the data is fully public and the bottleneck is purely analytical capacity. That makes it close to ideal for agent-based systems: no integration project, no data access negotiation, just expertise applied at a scale humans cannot match alone.
The judgment stayed human for the same reason it did in every engagement we run: the firm's product is findings it can defend, so agents propose and people confirm. The machines follow the money. The investigators make the case.