Banking AI agent · Stablecoin reserves
Stablecoin Reserve-Attestation AI Agent
No GENIUS Act obligation binds yet; the effective-date backstop is 18 January 2027. Build the reserve record and the examination pack now, because the CEO and CFO certify against whatever record exists then.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- CEO and CFO
- Build-ahead
- After launch
- Agent Care
What this agent does
Assembles the pack, not the certification
In
- 01
A reserve is held at a custodian, and the position is pulled from supported custody, ledger and chain sources.
- 02
A holding is tested against the closed asset list at GENIUS Act § 4(a)(1)(A), and anything outside it is named.
Reason
- 03
A Treasury position is re-tested for remaining maturity daily, because the limit is not a purchase-date test.
- 04
A reserve pool is tied to a single stablecoin brand, and a commingled position is flagged rather than allocated.
- 05
A coin count is reconciled from the mint-and-burn ledger to on-chain supply, and each difference is described.
Decide
- 06
A redemption settles, and the reserve movement behind it is matched to the published redemption policy.
- 07
A reserve use is logged against the § 4(a)(2) exception claimed for it, and an unmatched use stays open.
Out
- 08
A pack is retained as a dated version, alongside the composition disclosure and what each officer was shown.
- 09
A write action runs only inside the approval boundaries agreed during implementation.
Product statement
The agent maintains the record and assembles the pack; a registered public accounting firm examines it, and the CEO and CFO both certify.
Example workflow
One month, custody to certification
- 1Month-end positions receivedCustody statements, ledger extracts and on-chain supply reads
- 2Composition assembledEligibility, residual maturity, average tenor, custody location and coins outstanding
- 3Pack draftedComposition schedule, exception log, open items and confidence
- 4Controls appliedEligibility tests, maturity re-tests, brand-segregation checks and confidence threshold
No human action required
Stages 1 to 4 run unaided, and nothing is certified at any of them — the agent is assembling a record, and the treasury lane opens at the confidence gate.
Goes to the reserve owner to approve.
Adds a controller read first.
Reserve-owner approval
The pack is held with its custodian confirmations, its open items and the confidence.
Approve · Amend · Send to controller review
Approved — released to the accounting firm
- 6Reserve systems updatedOnly where write access and approval policy allow it
- 7Outcome evaluatedAmendment distance, open-item outcomes, examination findings and post-publication corrections
Amendments
Every reviewer amendment is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Certifying the monthly report — the CEO and the CFO, both.
- The monthly examination — a registered public accounting firm.
- Deciding that a reserve asset qualifies.
- Invoking an exception to the reuse prohibition.
Automation boundary
Agent acts unaided
- ✓Reconcile reserve positions across custodians to the ledger.
- ✓Re-test residual maturity and eligibility, day after day.
- ✓Tie coins outstanding to on-chain supply, and name the difference.
- ✓Draft the composition disclosure, and hold the pack for its owner.
Writes run only inside the boundaries agreed at implementation, and never ahead of approval.
- Publishing the monthly composition to the issuer's website.
- Setting or amending the public redemption policy.
- Engaging and instructing the accounting firm.
- Changes to eligibility, segregation or approval rules.
Example output
One reserve position, annotated
Everything the agent assembles is attached to the position it was drawn from.
Reserve record output · single position
Illustrative example- Custodian
- Primary reserve custodian
- Instrument
- Treasury bill · remaining maturity re-tested
- Remaining maturity
- 61 days
- Source of record
- Custodian statement
- Confidence
- 88%
- Eligibility basis
- Named item on the § 4(a)(1)(A) list
As received
Taken from the custodian statement and the month-end extract — nothing on this side is written by the agent.
Source facts used
- Custodian statement
- Month-end position file
- On-chain supply read
Why this is eligible
Remaining maturity is re-tested each day, so eligibility reads as a current fact.
What the score decides
Below the configured threshold the position picks up a controller read first.
Action
ApproveAmendSend to controller reviewValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every position
From the custody record
Test what the statute lists
Eligible assets are a closed list, and the limit runs on remaining maturity — GENIUS Act § 4(a)(1)(A).
Two officers, criminal exposure
A false certification imports the criminal penalties of 18 U.S.C. § 1350(c), by way of § 4(a)(3)(C).
Assembly, not attestation
The agent builds what the firm tests; the examination under § 4(a)(3)(A) is the accounting firm's, and its report stays the firm's.
Build an evidence trail
The reserve, the composition rule it was tested against and the officers who certified stay on the pack.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Custody and reserves
- BNY
- State Street
- Anchorage
- Fireblocks
Chain and token data
- Node and indexer APIs
- Mint-and-burn ledger
Treasury and general ledger
- Kyriba
- GTreasury
- Core banking
- GL extracts
Agent
Reserve record & examination pack
- Reads the positions
- Assembles the pack
- Holds for approval
Close and disclosure
- Workiva
- close tooling
- Website
- disclosure feed
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the pack
Six controls wrapped around one another. What none of them catches is named in the map beneath.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether the pack may stand
- L3
- Puts the release in the owner's hands
- L4 – L5
- Keep the reserve and the composition rule behind it
- L6
- Holds the pack at unreconciled when signals degrade
How Nestack evaluates it
Evaluate the assembly workflow — not only the finished pack.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the disclosure the public can read
- E1Final-output evaluationDid every reserve line trace to the custody position behind it?
- E2Step-level evaluationDid the agent use the right positions, eligibility rules and custodian configuration?
- E3Tool evaluationDid it read and write the correct custodian and the correct period?
- E4Confidence calibrationDo low-confidence positions actually attract more reviewer amendments?
- E5Slice evaluationHow does performance change across specific reserve classes?
- E6Business outcomeHow many lines were amended, or corrected after the disclosure was published?
Floor — the outcome the officers answer for
Failure modes
Where each failure originates in the agent
Seven failure modes, each placed at the stage it originates in.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- RG-03
Unconfirmed custodian balance
Prior month's position carried without a flag.
Stage gathers
Custody statements, chain data and ledger extracts
02 · Reasoning
2 modes
- RG-04
Maturity tested at purchase
Eligibility is not re-run as a bill ages.
- RG-06
Default exception label
A reserve use is logged under the wrong basis.
Stage proposes
Composition schedule, exception log and confidence
03 · Tool / write
2 modes
- RG-02
Coin supply unreconciled
Ledger count is not tied to on-chain supply.
- RG-05
Allocated, not segregated
One omnibus account is split by pro-rata share.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- RG-01
Disclosure differs from pack
Website figure is generated separately.
Stage returns
The pack the firm examines and the officers sign
05 · Change / Version
1 mode
- RG-07
Silent eligibility drift
A model or rule change alters what counts as eligible.
Stage tracks
Model, prompt, eligibility rules and custodian config
- Sev-1 · a wrong pack reaches a certification
- Sev-2 · a wrong line reaches the disclosure
- Sev-3 · source degrades, line goes to review
Affected slices
One reserve class can sit under a clean total
A programme-level reconciliation figure can read as clean while a few reserve classes absorb most of the amendments. Nestack reports the amendment rate by reserve class, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Tokenised money market funds | 8.6% | 3.7× | Review | |
| Omnibus custody accounts | 7.0% | 3.0× | Review | |
| Overnight repo positions | 4.0% | 1.7× | Watch | |
| Direct-held Treasury bills | 1.2% | 0.5× | Normal |
Bar: amendment-rate lift vs. direct-held-bill baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A cycle ends where a test begins
A cycle closes when the unreconciled reserve is a case the next release has to catch. That suite is what the next pack assembled is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Amendment rate rises in a reserve class.
- 02
Diagnose
The month the composition moved and the pack still said otherwise is read back until the cause narrows to one.
- 03
Improve
Number the correction, and the reserves that surfaced it stay fastened to it.
- 04
Verify
No release while a touched reserve case is outstanding; it reruns first.
- 05
Learn
The case is retained, and the composition rules change in the same commit.
Learn → Detect
The return edge. The next month is detected against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, reserve workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Reserve workflow discovery and boundary mapping.Week 1
- 02Custody, ledger and chain assessment.Week 1, Week 2
- 03Eligibility, maturity and segregation rule mapping.Week 1, Week 2
- 04Position ingestion and reconciliation.Week 2
- 05Pack assembly and source binding.Week 2, Week 3
- 06Confidence scoring and exception routing.Week 3
- 07Reserve-owner approval workflow.Week 3, Week 4
- 08Custody and ledger integration.Week 5
- 09Composition and segregation cases.Week 4, Week 5
- 10Guardrails and attestation controls.Week 4
- 11Reserve-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne brand, one month | ProductionProduction custody systems | AdvancedMultiple brands / issuers |
|---|---|---|---|
| Introduced at Pilot | |||
| Reserve record and reconciliation | ✓ | ✓ | ✓ |
| Reserve-owner approval | ✓ | ✓ | ✓ |
| Reconciliation-completeness baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by reserve class | — | ✓ | ✓ |
| Exception workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Custody and ledger integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-jurisdiction custody rules | — | — | ✓ |
| Multi-stage treasury approvals | — | — | ✓ |
| High reserve volume | — | — | ✓ |
| Multi-issuer reserve controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, reserve volume, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your custodian list and reserve policy
→Eligibility, maturity and segregation rule mapping
Week 102Representative prior month-end packs
→Position ingestion, reconciliation and the assembly baseline
Week 203Your documented eligible-asset definitions
→Reserve-policy mapping and automation-boundary definition
Week 104Access to relevant APIs, feeds or exports
→Custody, ledger and chain assessment, then integration setup
Week 205Packs you would not want examined
→Composition cases and the evaluation suite
Week 406What no certification may rest on
→Confidence scoring, exception routing, guardrails and approval controls
Week 307Named reserve owners to review the pack
→Reserve-owner approval workflow, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
The bands follow the monthly cycle rather than the slide, so the fifth week carries two phases.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Reserve workflow discovery, rule mapping and the automation boundary
W2Source integration and the reconciliation baseline
W3Assembly workflow, confidence logic and approval controls
W4Evaluation suite, eligibility checks and failure-mode testing
W5Custody integration, a pilot month and targeted corrections
W6One monthly cycle assembled under the chief financial officer, then handover
- Reading the band
- Each bar covers only the weeks its work is named in. The week 5 overlap is real, not padding.
- At the end of W6
- The last checks clear on a live cycle and Agent Care picks up monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Banking AI agent
Build the reserve record before the first certification falls due.
Show us your custodians, your eligible-asset policy and who signs. A wrong reserve line, once certified, is a criminal exposure carried by two named officers — not a supervisory finding.
- Typical delivery
- 4–6 weeks
- Deployment
- Your custody and chain data
- CEO and CFO
- Build-ahead
- After launch
- Agent Care available
Nestack Agents · Stablecoin reserve attestation
AGT-BK-18 · Agent Care available after launch