Book a demo
Topics · GOVERNANCE-AGENTS

Asset Mgmt Agent Identity Stop Control | Moring

Stop control is the ability to halt a running AI agent deliberately: pausing its actions mid-workflow, severing its credentials, and stopping delegated sub-agents, while everything else keeps running. It is distinct from decommissioning and broader than a kill switch, and it exists only where each agent holds a unique, verifiable identity.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Stop control is the ability to halt a running AI agent deliberately: pausing its actions mid-workflow, severing its credentials, and stopping delegated sub-agents, while everything else keeps running. It is distinct from decommissioning and broader than a kill switch, and it exists only where each agent holds a unique, verifiable identity.

How it works

Stop control is a distinct capability with three boundaries worth drawing. It is broader than a kill switch, which is the emergency variant, because stopping an agent is also a normal governance operation: pausing an agent whose behavior drifted, whose vendor integration failed, or whose purpose changed pending review. It is narrower than decommissioning, which ends the agent permanently. And it is an identity operation in every case, because halting one agent without halting the population requires that the agent holds credentials unique to it.

Why halting is hard by default follows from agent behavior. The CSA Agent Identity Governance Framework describes agents as acting at machine speed, spawning child agents, and persisting after the context that justified their privileges changes, so by the time a stop is decided, in-flight actions and delegated chains exist. Witness.ai's documentation of delegation chains across sub-agents and APIs that legacy identity models cannot track identifies precisely what a stop control must reach: the parent credential and every inherited authority beneath it.

The mechanics concentrate at the mediation layer. Lutril's MCP governance design gives the canonical form: the agent authenticates as its own identity, the gateway checks every tool call against policy, logs the call immutably, and can pause the agent within seconds, with a global kill switch as the backstop. IBM's agent gateway definition describes the same layer as a centralized control point mediating agent-to-tool and agent-to-agent communications, routing requests, managing permissions, and logging activity. A stop issued there covers every system the agent can reach.

What stop control is not clarifies what it must be. A practitioner thread on the IdentityManagement forum observes that identity programs have invested heavily in authentication, tight OAuth scopes, secret rotation, and short-lived tokens, while far less attention goes to what happens after authentication, when the agent is acting. Revoking a token stops future authenticated calls but not necessarily in-flight workflows or delegated chains; a complete stop control reaches the running state, which is why runtime monitoring, per Obsidian Security's identity-first controls, effective-authority mapping, and runtime monitoring, is part of the same design.

Stopping also has an evidence dimension that matters for an asset manager. The stop decision, its rationale, the actions halted, and the restoration steps are all retained, because a pause during review is itself a governance event. Dataiku's traceability finding, only five percent of more than 800 data leaders reporting AI output traceable all the time, applies to stops as well as outputs: an unrecorded stop is indistinguishable from a malfunction after the fact.

The regulatory frame treats stop control as part of proportionate oversight. SR 11-7 expects model risk governance commensurate with impact, and the ability to halt an agent acting on investment data is a control whose absence is a finding. SR 26-2 third-party guidance applies where the mediation layer or agent platform runs with vendors, including whether the firm can independently direct a stop. The EU AI Act, for firms operating in the European Union, expects human oversight provisions for higher-risk systems, and a working stop control is that oversight made operational.

Testing distinguishes control from intention, as with every halt capability. Periodic exercises pause a non-production agent mid-workflow, confirm in-flight delegated actions halt, confirm other agents continue, and confirm the retained evidence records the sequence. The asset management practice worth adopting is exercising the ordinary stop, not only the emergency one, because the ordinary stop is the one a review will actually need when behavior drifts or a purpose changes.

Frequently asked questions

How does stop control differ from a kill switch and from decommissioning?

Stop control is broader than a kill switch, which is the emergency variant, because stopping an agent is also routine governance: pausing an agent whose behavior drifted, whose integration failed, or whose purpose changed pending review. It is narrower than decommissioning, which ends the agent permanently with revocation, archive, and severed delegations. All three are identity operations requiring per-agent credentials, since halting one agent without halting the population is only possible when the agent holds credentials unique to it.

Why is stopping a running agent technically difficult without prior design?

Because agents act at machine speed, spawn child agents, and persist after their justification changes, per the CSA Agent Identity Governance Framework, so a stop decision arrives with in-flight actions and delegated chains already running. Witness.ai documents the delegation chains across sub-agents and APIs that legacy identity models cannot track, and the IdentityManagement practitioner thread notes programs invest heavily in authentication while neglecting what happens after, when the agent is acting. A complete stop must reach the running state, not just future authenticated calls.

Where is stop control implemented in an asset management architecture?

At the mediation layer between agents and the systems they reach. Lutril's MCP governance design shows the canonical form: the agent authenticates as its own identity, the gateway checks every tool call against policy, logs the call immutably, and can pause the agent within seconds, backed by a global kill switch. IBM's agent gateway description frames the same centralized point routing requests, managing permissions, and logging activity, so one stop covers every integration rather than requiring separate halts per system.

How does stop control serve an asset manager's oversight duties?

It operationalizes oversight: SR 11-7 expects model risk governance commensurate with impact, and the ability to halt an agent acting on investment data is a control whose absence is itself a finding. SR 26-2 third-party guidance covers vendor-run platforms, including whether the firm can independently direct a stop, and the EU AI Act expects human oversight provisions for higher-risk systems where firms operate in the European Union. The stop decision, rationale, halted actions, and restoration steps are retained, since Dataiku's traceability finding shows unrecorded events become indistinguishable from malfunctions.

Sources

  1. - federalreserve.gov
  2. - federalreserve.gov
  3. - artificialintelligenceact.eu