Contact
Platform Trust Intelligence Services About Contact

Agent Studio

Compose agents that carry a job from start to finish — and stop, by design, wherever a human must decide.

Agents that work while you sleep. Governance by design.

What an agent is here

In Agent Studio, an agent is a governed sequence of actions drawn from the firm's Action Catalog. Each action declares what data it may read, what it may propose, what it must cite, and which review tier its output requires. An agent is composed from those actions; it inherits every constraint they carry.

Fee Schedule Reconciliation

A worked example of a multi-level agent — from report to correction to approval.

  1. Run the report.

    The agent pulls the month's fee accruals and the current fee schedules for every account in its scope through the Bridge, read-only.

    Cites
    Account list, fee schedule versions, accrual extract.
    Approval
    None — reading within scope is pre-authorized.
    Recorded
    Scope, time, data hashes.

  2. Analyze the report.

    The agent compares each accrual to the governing schedule and the trust instrument's fee clause, and classifies every variance: rounding, tier boundary, missing minimum, wrong schedule, or unexplained.

    Cites
    Schedule line, instrument clause, prior-period treatment.
    Approval
    None.
    Recorded
    Every variance with its classification and evidence.

  3. Propose corrections.

    For each variance with a clear, cited cause, the agent drafts a correcting entry and a plain-language explanation. Unexplained variances are not corrected — they are escalated as exceptions.

    Cites
    The cause, the calculation, the correcting entry.
    Approval
    Required — routed by review tier.
    Recorded
    Proposal objects, untouched until reviewed.

  4. Human review.

    Clerical corrections go to the account administrator; structural changes (a schedule swap) to a trust officer; anything touching the instrument's interpretation to counsel or committee. The reviewer approves, edits, or rejects each proposal — the agent cannot skip, batch past, or time out this gate.

    Cites
    The proposal and its evidence, presented beside the source.
    Approval
    Required. The reviewer decides; the agent waits.
    Recorded
    Reviewer, decision, edits, timestamp, policy version.

  5. Apply the approved corrections.

    Only approved entries are written back through the Bridge, one action type at a time, with the approval reference attached.

    Cites
    The approval reference and the original proposal.
    Approval
    Already given — no further gate at write.
    Recorded
    The write, its acknowledgment from your system, and the hash-chained link to steps 1–4.

  6. Close the loop.

    The agent re-runs step 1 against the corrected data, confirms the variances are resolved, and files a summary to the reviewer. Anything still open stays open — visibly.

    Cites
    The post-correction extract and the original variance set.
    Approval
    None for the re-read; open items remain visible to the reviewer.
    Recorded
    Confirmation of resolved variances, and every item still open.

Six steps, four evidence types, one human gate — and a record that shows all of it in order.

Review tiers

Not every action carries the same fiduciary weight. Routing follows tier, with escalation when policy requires it.

Tier Example action Who approvesCan the agent write after approval
Clerical Rounding correction on a fee accrual One reviewer, sampled for qualityYes, with the approval reference
Structural A schedule swap Trust officer, single approval, full audit recordYes, with the approval reference
Interpretive Applying the instrument's fee clause to an ambiguous schedule Trust officer plus fiduciary counsel — two people, never oneYes, with the approval reference
Judgment Any action that will drive a distribution, a tax election, or a concentration position Committee action with written rationale on the recordYes, with the approval reference
  • Clerical Rounding correction on a fee accrual One reviewer, sampled for qualityCan the agent write after approval: Yes, with the approval reference
  • Structural A schedule swap Trust officer, single approval, full audit recordCan the agent write after approval: Yes, with the approval reference
  • Interpretive Applying the instrument's fee clause to an ambiguous schedule Trust officer plus fiduciary counsel — two people, never oneCan the agent write after approval: Yes, with the approval reference
  • Judgment Any action that will drive a distribution, a tax election, or a concentration position Committee action with written rationale on the recordCan the agent write after approval: Yes, with the approval reference

What agents can and cannot do

Can

  • Read within scope
  • Run and analyze reports
  • Draft corrections and memos
  • Escalate exceptions
  • Re-verify after approval
  • Run on a schedule or on an event

Cannot

  • Write without an approval reference
  • Act outside scope
  • Cite nothing
  • Overwrite a human decision
  • Give legal opinions
  • Authorize a distribution

Other agents you can compose

Overnight reconciliation

Compare positions, cash, and accruals against expected states and surface discrepancies before the team arrives.

Human gate: structural exceptions to the account officer.

Account exception surveillance

Watch for stale accounts, missing documentation, and elder financial exploitation patterns — anomalous distributions, changed payment instructions, unusual access.

Human gate: interpretive findings to officer and counsel.

Migration workflow automation

Guide conversion steps, validate mappings, and produce audit-ready progress reports under Parallel-Run Independence.

Human gate: cutover authorization by the named officer.

Governing-document term verification

Propose cited governing terms from the instrument against what the system of record holds, and flag disagreements.

Human gate: interpretive terms to counsel; judgment terms to committee.

Build your first agent with us.

We start with one report your team already runs by hand.

Request a demo