Fiduciary standards for autonomous systems.
Software cannot hold a duty. When an agent allocates, the fiduciary obligations — care, loyalty, prudence — remain exactly where they always were: with the humans and institutions that deployed it. The useful question is what the duty of care requires of a fiduciary who delegates work to a machine.
Delegation is not abdication
Fiduciary law has always permitted delegation — to sub-advisers, to index rules, to trading algorithms — under a consistent condition: the fiduciary must select prudently, instruct clearly, and monitor continuously. An autonomous agent is the newest delegate, not a new category of moral actor. Deploying one without the capacity to supervise it is the breach; the agent's sophistication is no defense.
What prudent deployment looks like
- Prudent selection: documented evaluation of the agent's validation record — the dossier described in the audit standards — before delegation.
- Clear instruction: the mandate expressed as machine-enforceable constraints, not aspirational prose. If a limit exists only in the IMA and not in the risk system, it is a wish.
- Continuous monitoring: drift and performance review at a cadence matched to the agent's autonomy — and a standing, rehearsed ability to halt it (kill criteria).
- Competence to oversee: a fiduciary who cannot understand the monitoring output cannot discharge the duty; oversight capacity is itself a staffing obligation.
Disclosure allocators should expect
Institutions allocating to AI-native strategies should treat the following as baseline diligence: the degree of autonomy in production (research assist vs. constrained execution vs. adaptive behavior); the independence of validation; the constraint register and where limits bind; incident history including every kill-criteria trigger; and the change-control process by which the agent's behavior can evolve. A manager unwilling to answer these has answered them.
The direction of travel
Accountability frameworks are converging on a simple architecture: machines execute; named humans remain answerable; evidence connects the two. The institutions that internalize this early will find that governance, far from slowing agentic adoption, is what makes it insurable, allocatable, and durable — the entire thesis of the QIS ecosystem.
Editorial synthesis of fiduciary principles under ERISA, the Investment Advisers Act, and trust law as applied to delegation and oversight. Not legal advice.