AGENT ARCHITECTURE

Not one model asked to do everything

Most AI in financial services fails for one reason: it guesses at numbers. Vident splits the work. Specialized agents handle the parts that need judgement and language. Everything with a correct answer is computed deterministically. A second pass validates the output, and a human approves it before anything touches a book.

In active development. Built from twenty years inside ETF portfolio management, trading and risk.

THE ROSTER

An agent per job, not a generalist

Each agent is scoped to a domain, given the tools that domain needs, and evaluated against that domain's own tests.

Portfolio

  • Position and holdings agents
  • Exposure, drift and attribution agents
  • Basket composition agents

Trading operations

  • Rebalance and order-preparation agents
  • Creation and redemption agents
  • Trade lifecycle agents

Reconciliation

  • Position and cash reconciliation agents
  • Corporate action and dividend agents
  • Exception triage agents

Research & documents

  • Market and research intelligence agents
  • Document retrieval agents
  • Institutional knowledge agents

THE DETERMINISTIC LAYER

Where there is a correct answer, it is computed

NAV, exposure, drift, basket composition and fee calculations have exactly one right answer, and a language model is the wrong tool for producing it. In Vident those values come from explicit calculations over source data, and the agent's job is to gather the inputs, choose the right calculation, and explain the result.

A language model never invents a number that lands in a report. When an agent needs a figure, it calls for it. The calculation is reproducible, reviewable, and can be recomputed by anyone checking the work afterwards.

ORCHESTRATION

A process that started this morning still knows where it was

Multi-step workflows run across systems and teams with persistent state and memory, so a workflow can pause for a data feed, a market event or a human decision and resume with full context.

Persistent state

Workflow state survives across steps, systems, and the gaps between them. Nothing is reconstructed from memory or from someone's inbox.

Tools and APIs

Agents act through explicit, permissioned tools rather than open-ended access. What an agent can reach is defined, not implied.

Retries and escalation

When a source is unavailable or an input is ambiguous, the workflow escalates to a human instead of guessing and moving on.

VALIDATION

Every material output gets a second, independent pass

Before anything reaches a person for approval, it is checked by a separate pass that did not produce it. The validator compares the output against source data, the originating instruction, and the firm's configured rules, and flags anything that does not reconcile.

The point is not that the first pass is untrustworthy. It is that a single unreviewed step is not an acceptable control in an operation where a wrong number has consequences.

THE CONTROL PLANE

Permissions, approvals, audit and evaluation

These are part of the architecture, not a compliance feature bolted on at the end.

Role-based permissions

What an agent can read, compute and propose is scoped by role. Access is defined, not inherited by accident.

Configured approvals

Approval chains are configured per workflow. Nothing touches a book without the right person signing off.

Complete audit logging

Every agent action and every human decision is recorded and attributable. If a regulator asks what happened, the answer is in the log.

Evaluation systems

Continuous measurement of whether the agents are still right after a model, prompt or data change, so quality is tracked rather than assumed.

INFRASTRUCTURE

Cloud or on-premises

SPECTER OS runs as a managed cloud deployment on AWS — Amazon Bedrock for model hosting, Step Functions for workflow orchestration, RDS for transactional state, S3 for documents and position files, Lambda for compute, and CloudWatch for logging and observability.

PRIVATE DEPLOYMENT

Runs inside your own environment

Most asset managers cannot send custody positions, holdings and counterparty exposure to a third-party model provider. For those firms, SPECTER OS runs entirely on their own premises — the same agent architecture, against models served locally on GPU hardware, with no data leaving the building.

Building for ETF operations?

SPECTER OS is being built from twenty years inside ETF portfolio management, trading and risk. If you run an ETF operation and want a say in how this develops, we would like to talk.

Start a conversation