Asset Management Registry Model Calls | Moring
The AI agent registry governs agent model calls in asset management by linking each agent to the validated model versions it may call, verifying at runtime that production calls match the approved versions, and recording call telemetry joined to the registry for lineage, cost, and audit purposes.
Short answer
The AI agent registry governs agent model calls in asset management by linking each agent to the validated model versions it may call, verifying at runtime that production calls match the approved versions, and recording call telemetry joined to the registry for lineage, cost, and audit purposes.
How it works
Agent model calls are where model risk meets agent governance in asset management. A firm can validate a model rigorously under SR 11-7, but if production agents call a different version, or call it through configurations the validation never contemplated, the validation is moot. The registry is the control that keeps what is validated and what is called the same thing.
The linkage is structural. Each registry entry records the model version or versions the agent may call, with the validation reference and approval date. Databricks' Unity Catalog pattern demonstrates the mature architecture: models, agents, and data governed under one framework with lineage and inference traces, so the connection between an agent and its model is a first-class governed relationship rather than a configuration detail buried in code.
Runtime verification closes the loop. The enforcement layer checks, at call time, that the agent is invoking a registered, validated version; calls to unregistered or superseded versions raise conflicts routed to the model risk function. Adobe's analysis of model governance shows the failure mode this prevents: teams retraining models without approval, other teams unable to verify which model generated an output, and version drift accumulating silently until an audit or a bad outcome exposes it.
Telemetry joined to the registry turns model calls into an observable surface. Inference traces show which agent called which model version with what inputs, feeding lineage for audit, cost attribution for the economics of model selection, and drift detection when call patterns diverge from the recorded configuration. Snowflake's AI governance guidance frames the broader need: enterprises must control what models and agents exist, what data each can touch, and how it is proven, and model call records are the proof in question.
Model call oversight also disciplines model change. When a model version is retired or superseded, the registry identifies every agent calling it, so migration and revalidation happen estate-wide rather than agent by agent. This is the SR 11-7 expectation that model changes be governed and documented, applied to the agent layer where changes previously escaped notice.
Cost enters the picture through the same record. Model calls carry token consumption, and the registry's linkage lets the firm compute cost per model version per workflow, informing the model selection decisions that validation quality thresholds should govern. A cheaper model is admissible only where the validated quality bar is met, and the registry carries both the version and its validation reference that make that determination possible.
The result is an estate where the firm can answer the foundational model risk question for any agent: which model produced this output, was that version validated, under what approval, and with what call history. That answer, available from the registry rather than reconstructed across teams, is what lets asset management firms extend agentic automation into portfolio-relevant workflows while the firm's model risk framework and its fiduciary obligations remain demonstrably in force.
Frequently asked questions
How does the registry govern which models an agent may call?
Each registry entry records the model versions the agent is approved to call, with validation references and approval dates, and the enforcement layer verifies at call time that production calls match the registered versions. Calls to unregistered or superseded versions raise conflicts routed to the model risk function, keeping what was validated and what is actually invoked aligned across the estate.
Why is model call verification important for asset management?
Because agents make models actionable: a rigorously validated model that reaches production through drifting agent calls undermines the validation work performed. Runtime verification, supported by unified lineage across agents, models, and data, prevents the silent version drift that model governance research identifies, where teams retrain without approval and outputs cannot be traced to a verified model version.
How does model call telemetry support oversight and economics?
Joined to registry entries, inference traces show which agent called which model version with what inputs, supporting audit lineage, drift detection, and cost attribution per model version per workflow. Cost per call informs model selection, with cheaper admissible only where validated quality thresholds are met, and the registry carries both the version and validation reference that make that determination evidence-based.
What happens when a model version is retired or superseded?
The registry identifies every agent calling that version, so migration and revalidation proceed estate-wide rather than agent by agent, and call verification blocks the retired version in production until each agent is migrated. This extends the model change governance that model risk frameworks require to the agent layer, where version changes previously escaped notice until audits or bad outcomes exposed them.