Models that read chain state, and a human who signs.
AI agents for blockchain read on-chain state continuously and turn it into evidenced proposals. neoAgent™ covers collateral valuation, compliance and sanctions screening, treasury drift monitoring, and transaction risk scoring. Every output carries its inputs, model version, confidence, and rationale. A human with dual control signs.
neoAgent · collateral valuation
runningthe agent proposes and evidences; it never signs
24/7
Continuous on-chain monitoring
0
Signing authority held by agents
100%
Decisions replayable
3.2s
Observe to proposal, median
Four jobs that are too frequent for a person and too contextual for a rule
The AI work that prices markets and segments players applies directly to on-chain state. Here the record is public, the inputs are verifiable, and the output must be reproducible.
Valuation & NAV attestation
Collateral and tokenized asset values recomputed continuously from oracle feeds, registries, and custodian data, with a confidence figure and the inputs attached to every attestation.
Compliance screening
Counterparty addresses assessed against sanctions lists, chain-analytics risk, and your own policy before broadcast, with the reasoning written to the audit trail.
Treasury monitoring
Tier ratios, reserve coverage, and exposure drift watched continuously against policy, so a rebalance is proposed when the position moves.
Transaction risk scoring
Withdrawal and transfer patterns scored against the entity's own baseline, flagging the sequence that does not fit: structuring, sudden counterparty changes, dormant-account reactivation.
What an agent may do, and what only a person may do
This table is the constraint that makes the capability safe to deploy against assets. It is not a configuration option, and no tier changes it.
| Action | Authority | Detail |
|---|---|---|
| Read chain state | Agent | Continuous, no human in the path |
| Compute a valuation | Agent | With inputs and confidence recorded |
| Draft an attestation | Agent | Proposed, never published unilaterally |
| Flag a counterparty | Agent | Advisory, routed to a reviewer |
| Hold a transaction | Agent | Fail-safe: holding is always permitted |
| Release a held transaction | Human | Never automated, in any configuration |
| Sign or broadcast | Human | Signing authority is never delegated to a model |
| Move treasury funds | Human | Dual control, unchanged by the presence of agents |
A decision nobody can reconstruct is not auditable
Supervisors ask three things about an automated decision: what data it saw, what produced the output, and who was accountable. All three are recorded on every decision at the time it is made.
Every decision is evidenced
An agent output carries the inputs it read, the model and version that produced it, the confidence, and a plain-language rationale, retained and queryable.
Deterministic replay
Any past decision can be replayed against the exact inputs and model version that produced it. What a regulator asks about an automated decision is how it was reached, and the answer must be reproducible.
Fail-safe by construction
Every failure mode holds. A degraded oracle, a low-confidence output, or an unreachable model results in a transaction waiting for a human, never in one going through unchecked.
Recorded on every agent decision
Proposals arrive where your team already works
Agents read chain state directly and take feeds over standard interfaces, so they run against your own custody, registry, and treasury systems. Proposals arrive in the tooling your team already uses.
- Runs against your existing custody and registry systems
- Proposals pushed into your case-management or approval tooling
- Every decision replayable against pinned inputs and model version
// 200 OK{ "asset_id": "ast_2C91AF", "nav_eur": 4213880.42, "confidence": 0.94, "inputs": [ { "source": "chainlink:eur-usd", "age_s": 41 }, { "source": "registry:land-de", "age_s": 86400 }, { "source": "custodian:report", "age_s": 3600 } ], "model": "neoagent-valuation-3.2", "status": "proposed", "signed_by": null}What risk and compliance teams ask
On-chain state moves continuously, and the alternative is a nightly export read the next morning. Recomputing a valuation when an oracle updates, noticing a counterparty change, and flagging a pattern across thousands of transfers are too frequent for a person and too contextual for a rule.
See one of your manual decisions with the evidence attached
A working session with the engineers who build these: which of your on-chain decisions an agent could surface with evidence attached, where the authority boundary would sit, and what a replayable decision record looks like.