Build your own agents
Build the agent your policy already describes.
Studio is a visual workflow canvas for composing agents against your own policy sets and knowledge. You set the guardrails and the human-review gates, version each change, evaluate against labelled test cases, and publish only when the evaluation passes.
How it works
Composed, evaluated, then published — in that order.
An agent that touches regulated work should be built the way a control is built: written down, tested against known cases, and released only when it passes. Studio makes that the only available path to production.
- 01
Compose on the canvas
Steps, tools, policy sets and knowledge are assembled on a visual workflow canvas, so a reviewer who does not write code can read what the agent will do.
- 02
Set guardrails and gates
Guardrails constrain what the agent may touch. Human-review gates mark the points where a person must accept the work before it continues.
- 03
Evaluate, then publish
Each version is evaluated against labelled test cases. Publish is gated on the result: a version that fails its suite cannot be released.
What it produces
A governed agent with a version history.
What leaves Studio is not a prompt in someone's notebook. It is a versioned, owned, evaluated service that Control can register and audit.
Visual workflow canvas
Compose the agent as a readable workflow rather than as configuration, so second-line review can be done on the design itself.
Your policy sets and knowledge
Agents are built against the policies and documents your institution already maintains, under the same entitlements as the people who use them.
Guardrails
Explicit limits on the tools, data and actions available to the agent, enforced at run time rather than requested in a prompt.
Human-review gates
Named points where a person must accept the work before it proceeds, so an agent cannot complete a chain that policy says a human must break.
Versioning
Every change is a version with an owner and a history. Releases roll back in one step.
Evaluation against labelled cases
Suites of labelled test cases run against each version, and publish is blocked when a suite falls below its threshold.
Where it stops
Studio cannot build past your controls.
An agent composed in Studio inherits the platform's constraints rather than escaping them: it holds its own least-privilege identity, cannot call an unapproved model, and cannot take an action that requires an entitlement it does not have. Removing a review gate is itself a governed, logged change.
An agent that fails its evaluation suite does not ship. That gate is not a setting a builder can turn off to meet a deadline.
Works with the rest of the platform
One record, one permission model, one audit trail.
Compliance
Alert triage, adverse media and a case narrative drafted for the MLRO — never filed without one.
Exceptions and service levelsOperations
Exception queues with live SLAs, and release packs that stop at the authoriser rather than past them.
Governance and auditControl
Model registry, agent registry, entitlements and an immutable event log covering every human and agent action.