The Governance Layer · Framework

The rules for trusting an autonomous agent with capital do not yet exist.

Autonomous agents can now research, construct, and execute financial strategy end to end. The frameworks required to trust them at institutional scale — authority, oversight, auditability, accountability — remain unwritten. Existing regulation governs models, disclosures, and conflicts of interest. None of it squarely governs an autonomous agent taking an action with capital. This is the map of what must be governed before it can.

↓ The eight questions ↓ Where the rules stand

The control loop every autonomous financial system implies. Each existing regulation governs one or two of these stages. None governs the loop as a whole.

The Framework

Eight questions autonomous finance must answer.

Before an institution can delegate capital to an autonomous agent, eight questions must have answers. Today, most do not — not because the answers are hard, but because no authority has yet been charged with writing them. Each question below is stated as a canonical definition, then mapped to what current regulation does and does not address.

The question of scope

Authority

Authority is the boundary between the decisions an autonomous system may take on its own and those that require human authorization.

Today

Model-risk guidance governs whether a model is sound; securities law governs disclosures and conflicts. Neither defines the scope of action an agent is permitted to take.

The gap

No framework draws the line between decision-support and autonomous execution — the exact line that matters the moment an agent can place a trade.

The question of intervention

Oversight

Oversight is the standard governing when a human must be able to observe, pause, or override an autonomous system — and how the system escalates to them.

Today

The EU AI Act requires human oversight for high-risk systems (Article 14); the NIST framework includes a Manage function. Both assume a human supervising a tool.

The gap

Neither specifies escalation thresholds or halt criteria for an agent acting continuously and unsupervised in live markets.

The question of change

Drift

Drift is the divergence of a system's live behavior from the baseline that was validated before deployment.

Today

The revised US model-risk guidance (SR 26-2) requires ongoing monitoring — but the validate-then-deploy discipline assumes a largely stable system.

The gap

Agents that adapt continuously break the assumption that the validated model is still the model that runs. Continuous learning has no continuous-validation standard.

The question of record

Auditability

Auditability is the property that every autonomous decision can be reconstructed, explained, and attributed after the fact.

Today

Books-and-records rules and logging guidance establish that records must exist and be retained.

The gap

There is no standard for what constitutes a sufficient decision record for an autonomous agent — the equivalent of the trade ticket and rationale a human leaves behind.

The question of responsibility

Accountability

Accountability is the assignment of fiduciary and legal responsibility for an outcome that no human directly decided.

Today

Fiduciary duty under the Investment Advisers Act attaches to the adviser; the SEC applies existing fiduciary and antifraud law to AI use.

The gap

Fiduciary law presumes a human decision-maker. It is unresolved whose duty is breached when the decision was, in substance, the agent's.

The question of provenance

Identity

Identity is the verified provenance, authority, and integrity of an autonomous agent that a counterparty must trust in order to transact with it.

Today

NIST's AI Agent Standards Initiative, announced February 2026, is developing identity and authorization guidance, with a profile targeted for late 2026.

The gap

No operational standard yet exists for one institution to verify another's agent — its identity, its mandate, and that it is what it claims to be — inside a live counterparty relationship.

The question of scale

Systemic behavior

Systemic behavior is the market-level effect of many autonomous systems acting on correlated signals simultaneously.

Today

Market-structure rules address speed, outages, and circuit breakers — the mechanics of fast markets.

The gap

No framework addresses correlated autonomous decisioning: the herding and feedback risk of many agents allocating capital on the same read at the same instant.

The question of common ground

Standards & interoperability

Standards and interoperability are the shared definitions, protocols, and profiles that let autonomous systems from different institutions operate together and be governed by a common reference.

Today

NIST's AI Agent Standards Initiative is developing interoperability guidance, with an Agent Interoperability Profile targeted for late 2026; ISO/IEC 42001 offers a management-system standard the NIST framework crosswalks to.

The gap

There is no settled common standard for how autonomous financial agents identify, authorize, and transact with one another across institutional boundaries. Private conventions do not compose into a market.

The Evidence

Where the rules stand — mid-2026.

The gap is not an opinion. It is visible in the current state of the four frameworks that come closest. Each governs part of the control loop. None was written for an agent that acts on its own.

US Model Risk

SR 26-2 · Fed / OCC / FDIC · Apr 2026

On 17 April 2026 the US banking agencies replaced the 2011 model-risk guidance (SR 11-7) with revised, risk-based guidance. It carries forward the core disciplines — sound development, independent challenge, ongoing monitoring — but industry analysis reads generative and agentic AI as flagged for separate, later treatment rather than squarely covered. It governs whether a model is sound; it does not govern an agent's authority to act. Primary source →

EU AI Act

Reg (EU) 2024/1689 · in force Aug 2024

The Act treats certain financial uses — credit scoring, insurance pricing — as high-risk, imposing risk-management, data-governance, and human-oversight obligations. But as of mid-2026 those high-risk obligations were provisionally deferred toward late 2027, and autonomous trade execution is not squarely named among the high-risk uses. It governs financial decisioning about people, not an agent deploying capital. Official timeline →

NIST AI RMF

Voluntary · Agent Initiative Feb 2026

The NIST framework (Govern, Map, Measure, Manage) and its Generative AI Profile are the most-referenced voluntary standard, but they describe a tool under human management. In February 2026 NIST launched an AI Agent Standards Initiative that explicitly acknowledges the agent governance gap, with agent-specific guidance only planned for late 2026. The leading standards body is, in effect, on record that this space is unwritten. NIST AI RMF →

US Securities

SEC · no AI-specific rule

The SEC has no AI-specific rule on the books. Its 2023 Predictive Data Analytics proposal — the one rule that would have directly constrained AI in investor-facing decisions — was withdrawn in June 2025. The agency now governs AI through existing antifraud, fiduciary, and disclosure law, and through enforcement against overstated AI claims. Existing law reaches the marketing of AI far more clearly than the autonomy of it.

The Position

Why the map must come before the rules.

QISTrust position

The sections above are a map, not an argument: each is anchored to a public, current source, and states what exists and what does not. What follows is QISTrust's position, offered as such.

Fields do not become governable by waiting for regulation. They become governable when someone states the questions precisely enough that answers — regulatory or contractual — can be written against them. Model risk had SR 11-7 for fifteen years before this year's revision; cybersecurity had the NIST framework long before it had law. In each case, a structured articulation of the problem came first, and the rules were measured against it.

Autonomous finance has no such articulation yet. Our position is that governance for autonomous systems must be designed as infrastructure, built in before deployment — not retrofitted as compliance after an incident. The eight questions are the beginning of that articulation. We will be wrong about some of them, and we will revise in the open. But the field is served better by a precise map that can be corrected than by silence that cannot.

This page is informational and educational. Nothing here is legal, investment, or compliance advice.

Across the ecosystem

The loop, distributed.

Governance is one stage of a larger system. The agents that act, and the capital they move, are documented on the other two layers.

Agent · IntelligenceThe systems that research, construct, and execute — the action this framework governs. Fund · CapitalThe strategy taxonomy: what autonomous systems are being trusted to allocate. Trust · GovernanceThe Agent Oversight Framework: the lifecycle controls that operationalize these questions.

Sources

The evidence base.

Regulatory status is stated as of July 2026 and is fast-moving; each anchor links to a primary or official source. Where this page takes a position, it is labeled as a position.

Federal Reserve / OCC / FDIC — SR 26-2, Revised Guidance on Model Risk Management (April 2026), superseding SR 11-7 (2011). European Union — AI Act (Regulation (EU) 2024/1689) implementation timeline, European Commission AI Act Service Desk. NIST — AI Risk Management Framework, Generative AI Profile (AI 600-1), and the AI Agent Standards Initiative (CAISI, 2026). US SEC — withdrawal of the Predictive Data Analytics proposal (June 2025); regulation of AI through existing antifraud, fiduciary, and disclosure frameworks.