Asset Management Agent Governance Version Control | Moring
Version control in AI agent governance means treating agents, their prompts, their tool permissions, and the governing policies themselves as versioned artifacts with recorded changes. For asset managers, unversioned agents are ungovernable: nobody can say which policy authorized an action or which configuration produced an output. Versioned change control extends the firm's existing model change discipline to every layer agents touch.
Short answer
Version control in AI agent governance means treating agents, their prompts, their tool permissions, and the governing policies themselves as versioned artifacts with recorded changes. For asset managers, unversioned agents are ungovernable: nobody can say which policy authorized an action or which configuration produced an output. Versioned change control extends the firm's existing model change discipline to every layer agents touch.
How it works
Version control is the quiet foundation of agent governance. An agent is not a static application: its prompts change, its tool permissions change, the models behind it are updated by vendors, and the policies that constrain it evolve. If none of those layers are versioned, an asset manager cannot answer a basic governance question after an incident or an audit request: which configuration, under which policy, produced this action on this date. Research on agent governance frames the challenge in terms of behavior and access, and versioning is what makes both reconstructable.
The two-layer structure described in enterprise governance research sharpens the point. Agent governance has two layers: controlling what agents can do and governing what data they can consume, and most frameworks govern agent behavior while leaving the data layer unmanaged, where analysis finds the majority of AI failures originate. Version control applies to both layers: the agent's own artifacts, prompts, tools, and permissions, and the data policy versions that define what it may read. When an asset manager changes a data entitlement for an agent that touches performance data, that change should be as recorded and reviewable as the change to the agent itself.
What should be versioned, concretely: the agent definition, including its instructions and prompts; its tool and API permission set; the policy file that governs it, ideally stored and diffed as code; the model version it runs on, with vendor updates tracked as events rather than absorbed silently; and the risk classification that determined its control tier. Each change should carry an author, a reviewer, and a rationale. This mirrors change management the firm already applies elsewhere, and asset managers have deep precedent: portfolio mandates, valuation methodology, and client reporting templates all change under documented control.
The regulatory rationale is direct. Model risk management under SR 11-7 and its successor SR 26-2 requires asset managers to document model changes, validate them independently, and retain the evidence; agents that influence valuations, client reporting, or operations deserve the same discipline because their risk character is comparable. A versioned history lets validation teams review exactly what changed between agent versions rather than reviewing the whole system each time, and it gives internal audit a sampling frame: pick a date, retrieve the configuration and policy in force, and verify observed behavior against them.
Version control also disciplines vendor-side change. Cloud governance guidance emphasizes that agents operate with delegated authority and affect multiple systems at once, and much of that authority is exercised through components the firm does not build: foundation models, tool libraries, and platform runtimes updated on vendor schedules. Treating vendor model updates as version events, with regression checks on agent behavior before the new version serves production traffic, prevents silent behavioral drift from becoming a governance incident. Security researchers make the same point about agent sprawl: capabilities and environments change faster than annual review cycles, so the version record is the only reliable memory.
Operationally, an asset manager should keep policy and agent configuration in the same versioned repository as the code they constrain, so a single change history covers behavior, permissions, and enforcement. Promotion between environments, development, pre-production, production, should be an explicit, reviewable event, with the governance gate checking that the version being promoted matches the version that was classified and approved. Rollback, too, should be a first-class capability: when an agent misbehaves, restoring the prior version and policy is the fastest control, and it only works if the prior state is intact.
The payoff is examinability. When every layer is versioned, the answer to which policy authorized this action becomes a query against recorded state rather than a search through chat threads and tribal memory. For an asset manager facing model risk expectations under SR 26-2, and EU AI Act deployer duties where European operations apply, that queryability is the difference between demonstrating control and narrating it. Version control costs little, compounds steadily, and underwrites every other governance mechanism the firm deploys.
Frequently asked questions
Which agent artifacts need version control in an asset management governance program?
Five layers at minimum: the agent definition including its instructions and prompts; its tool and API permission set; the governing policy file, ideally stored and diffed as code; the model version it runs on, with vendor updates tracked as explicit events; and the risk classification that set its control tier. Governance research distinguishes the behavior layer from the data layer, so entitlements governing what an agent may read should be versioned with the same rigor, each change carrying an author, a reviewer, and a rationale.
How does version control support the model risk obligations an asset manager already has?
It gives those obligations a substrate. Model risk management under SR 11-7 and continued in SR 26-2 requires documented changes, independent validation, and retained evidence for systems influencing decisions. Versioned agent configurations let validation teams review precisely what changed between agent versions instead of re-reviewing whole systems, and they give internal audit a clean sampling frame: pick a date, retrieve the configuration and policy in force, and verify observed behavior against them. Without versioning, none of that reconstruction is possible.
How should an asset manager handle vendor model updates that change agent behavior?
Treat each vendor model update as a version event rather than an invisible change. Track which model version each agent runs on, run regression checks on agent behavior before the new version serves production traffic, and record the review decision alongside the version history. This matters because agents exercise delegated authority across multiple systems, and silent behavioral drift from vendor-side updates can undermine controls the firm believes are in place, surfacing later as unexplained changes in outputs or tool usage.
What is the connection between version control and rollback capability?
Rollback depends entirely on versioning. When an agent misbehaves, restoring the prior version of its configuration, permissions, and governing policy is the fastest and most reliable control, and it works only if the prior state was recorded intact. Governance should therefore make rollback a first-class, reviewable event rather than an emergency improvisation, with promotion and demotion between environments explicit. An asset manager that can demonstrate controlled rollback also demonstrates the intervention capability that examiners and deployer obligations increasingly expect.