Article contents
Regulatory information can change. This resource is educational and should not be treated as legal, clinical, or regulatory advice. Consult the applicable primary guidance and qualified professionals for decisions affecting your organization. Last reviewed November 18, 2026.
There is no single rulebook1
Outside the EU, AI rules are arriving piece by piece. In the United States there is still no federal AI law that covers businesses in general. An executive order signed on 11 December 2025 directs federal agencies to challenge state AI laws the administration considers burdensome, and names Colorado's law as an example. An executive order does not repeal a state law, though, so state rules stay on the books unless a court or Congress says otherwise.
Meanwhile federal agencies, standards bodies and other countries have published their own expectations. Few of them mention agents by name. All of them describe what happens when software makes or carries out decisions that affect people.
What changed, place by place2,3,4,5,6,7,8,9,10,12,13
Colorado
The 2024 Colorado AI Act never took effect on its original date. SB 26-189, signed in May 2026, repeals and reenacts it for automated decision-making technology used in consequential decisions such as hiring, lending, housing, insurance and health care, made on or after 1 January 2027. Deployers must give notice, offer an easy way to ask for more information after an adverse outcome, let people correct their data and request meaningful human review, and respond within 30 days.
California
Regulations from the California Privacy Protection Agency took effect on 1 January 2026. Businesses that use automated decisionmaking technology for significant decisions about consumers have until 1 January 2027 to meet its requirements, and risk assessments and cybersecurity audits follow on their own timelines.
US federal agencies
OMB memo M-25-21 gave agencies 365 days to put minimum risk practices in place for high-impact AI, including human oversight, intervention and accountability, a fail-safe where practicable, and timely human review or appeal for people affected by AI-enabled decisions. Vendors selling agents to government inherit these expectations.
NIST
In February 2026 NIST launched the AI Agent Standards Initiative, and its cybersecurity center published a concept paper on identity and authorization for software and AI agents. More than 600 commenters responded; the first demonstration will cover agents in software development. It is guidance, not law, but it shows where standards are heading.
Singapore
IMDA's Model AI Governance Framework for Agentic AI, first published in January 2026 and updated in May, asks organizations to bound risks up front by limiting autonomy, tools and data access, to make humans meaningfully accountable with approvals at significant checkpoints, and to audit whether those approvals still work.
South Korea
The AI Basic Act and its enforcement decree took effect on 22 January 2026. Operators must assess whether a system is high-impact AI, a list that includes hiring and loan screening, before deploying it, and notify users in advance when generative or high-impact AI is used. The government announced a grace period of at least one year.
The same asks in different words
Our summary of the overlap. The exact duties, thresholds and dates differ, so check each source for the rules that apply to you.
Why agents raise the stakes
Most of these rules were written with scoring and recommendation systems in mind: a model produces a score and a person decides. Agents also act. When an agent declines a refund, rejects an applicant or changes a price itself, the decision and the action are the same step.
That changes where oversight has to sit. A person who reads a report next week cannot provide the human review these rules describe. The review has to happen before the action, or the action has to be reversible and visible to someone who can reverse it. And the record has to show which agent acted, under whose authority, with which permissions, at the moment it acted.
Identity is becoming the foundation6,7,11
NIST's agent work makes a point that is easy to miss: many early agent deployments repeat old identity mistakes, such as shared credentials, broad standing access and actions that cannot be traced to a specific actor. Its guidance is to treat agents as identities in their own right, with authentication, narrowly scoped authorization and accountability tied back to the people responsible.
Singapore's framework says the same in operational terms: decide up front what each agent may access and do, roll out gradually, and keep monitoring. Neither is legally binding. Both are the kind of guidance regulators, auditors and enterprise customers will measure you against.
Build once, map to many
These controls line up with most of the rules above. Have a qualified adviser confirm what applies to your organization.
- One inventory of agents with owner, purpose, tools and the people each one affects
- A documented classification of each agent's use, updated when its tools change
- A separate identity for every agent, never a shared human account
- Permissions enforced by the platform, scoped to the tools and data each agent needs
- Approval before actions that affect people, money or anything hard to undo
- A way for affected people to ask for an explanation and a human review
- A clear AI notice wherever an agent talks to people
- A stop and pause that work at once, whatever the agent is doing
- A complete record of runs, tool calls, approvals and changes, kept and exportable
- A regular review of whether approvals still catch problems
Where Quaspar fits
Quaspar is the governance layer for AI agents. It is built around the controls on this list: each agent's permissions are set when it is created and enforced by Quaspar rather than the model; outgoing messages, submissions and changes to connected systems wait in Workspace for a person; an administrator can cancel any run or pause any agent instantly; and every agent run, tool call, approval, rejection and permission change is recorded with who, what, when and the outcome, ready to export.
We do not offer legal advice or certify compliance with any law, and Quaspar is not SOC 2 certified. What we offer is one place where the people responsible for your agents can see and control what those agents do. Quaspar is in private preview.
Next step
Quaspar is the governance layer for AI agents and is in private preview. Request access to set agent permissions, route important actions to people and keep a record of what every agent did.
Request Studio accessSources
Authoritative references
Rules, standards, and guidance can change. The links below were reviewed on November 18, 2026.
- 1The White House. Ensuring a National Policy Framework for Artificial Intelligence (11 December 2025)
- 2Colorado Attorney General. Automated Decision-Making Technology & Chatbot Safety Rulemaking
- 3Colorado Legislative Council Staff. 2026 Summary of Legislation: Information Technology and Artificial Intelligence
- 4California Privacy Protection Agency. California Finalizes Regulations to Strengthen Consumers' Privacy
- 5Office of Management and Budget. M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust
- 6National Institute of Standards and Technology. Announcing the AI Agent Standards Initiative
- 7NIST National Cybersecurity Center of Excellence. Accelerating the Adoption of Software and AI Agent Identity and Authorization (concept paper)
- 8Infocomm Media Development Authority, Singapore. Model AI Governance Framework for Agentic AI, version 1.5
- 9Infocomm Media Development Authority, Singapore. Factsheet: Model AI Governance Framework for Agentic AI
- 10Ministry of Science and ICT, Republic of Korea. The AI Basic Act Comes into Force
- 11National Institute of Standards and Technology. Back to the Future: Why Agentic AI Needs a Strong Identity Foundation
- 12U.S. International Trade Administration. South Korea AI Basic Act
- 13National Institute of Standards and Technology. Comments on Software and Agentic AI Identity Concept Paper
