---
title: What does FedRAMP mean for AI agents used by US federal agencies?
url: https://opencontroller.lyzr.com/blog/fedramp-and-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What does FedRAMP mean for AI agents used by US federal agencies?

**FedRAMP authorises cloud service offerings, not individual AI agents. If an agent runs on, or sends federal data to, a commercial cloud service, that service generally needs a FedRAMP authorisation at the right impact level, and the agency still needs its own authorisation to operate (ATO) for the system that uses the agent. Agents add new things to evidence inside that boundary: which agents exist, what they can access, and what they did.**

Federal teams ask two questions: "Is the model provider authorised?" and "What else do we need for agents?" The first is about the vendor. The second is about you.

## What does FedRAMP actually cover?

FedRAMP, run by the US General Services Administration (GSA), is the government-wide programme for security assessment and authorisation of cloud products. Its unit is the cloud service offering and its authorisation boundary. Impact levels follow FIPS 199 (Low, Moderate, High).

For AI, that means:

- **The model API or AI platform** is a cloud service. Its authorisation covers what is inside its boundary.
- **Your agent** is usually software you build or configure on top. It inherits controls from the authorised service but adds its own risk: prompts, tools, credentials, data flows.
- **Every SaaS the agent calls** (CRM, ticketing, knowledge tools, MCP servers) is a separate boundary question.

## What was the FedRAMP AI prioritisation?

In August 2025 FedRAMP launched an initiative to prioritise authorisation of AI cloud services providing conversational AI engines for routine use by federal workers. According to [FedRAMP's AI page](https://www.fedramp.gov/ai/), eligible services had to meet five criteria, including enterprise features (SSO, SCIM provisioning, role-based access control, real-time analytics), data separation, demand from at least five CFO Act agencies or a CIO Council recommendation, availability through the GSA Multiple Award Schedule, and the ability to meet FedRAMP 20x requirements within two months. The initiative completed in April 2026 and is closed to new entrants.

Note what it did not cover: it was scoped to conversational AI engines. It did not authorise the agents agencies build on them.

## What is FedRAMP 20x and why does it matter for agents?

FedRAMP 20x is GSA's effort to move authorisation from narrative documents to automated, machine-readable evidence measured through Key Security Indicators. It has run as phased pilots, starting with Low and then Moderate. GSA's published roadmap also signals an end to new Rev. 5 applications; check [fedramp.gov](https://www.fedramp.gov/) for current dates, as they have shifted.

For agents, the direction matters more than the dates. Continuous, machine-readable evidence suits agent estates, where configurations, tools and permissions change weekly. A quarterly screenshot of an agent's permissions is stale before the review ends.

## What evidence do agencies need for AI agents inside an authorised boundary?

Map agent-specific evidence to existing NIST SP 800-53 control families. A starting point:

| Evidence item | Typical 800-53 family | What it shows |
|---|---|---|
| Inventory of agents, versions, models, tools and MCP servers | CM (Configuration Management) | What exists and what changed |
| Agent identities and delegated on-behalf-of scopes | IA, AC | Who the agent acts as and what it can reach |
| Risky permission and configuration findings | RA, CA | Continuous assessment of agent posture |
| Prompt, tool-call and model-call logs | AU | What the agent did, traceable to a user |
| Guardrail and policy decisions (allow, redact, block) | SI, AC | Controls operating in the request path |
| Kill and quarantine actions with timestamps | IR | Response capability when an agent misbehaves |
| Evaluation and red-team results per version | SA, RA | Pre-deployment and ongoing testing |

This is evidence for controls, not a claim of authorisation. Your assessor decides adequacy. For overlays aimed at AI, see [NIST control overlays for AI agents](/blog/nist-control-overlays-for-ai-agents). Agencies also have OMB obligations for AI use; see [OMB M-25-21 and federal AI agents](/blog/omb-m-25-21-and-federal-ai-agents).

## Checklist for an agency deploying agents

1. Confirm each model provider and platform has a FedRAMP authorisation at the needed impact level, and that the specific service and region you use is in its boundary.
2. List every downstream system each agent can call, and check each one's authorisation status.
3. Register every agent with an owner, purpose, model, tools and data classification.
4. Give agents distinct identities; do not reuse a human's credentials or a shared key.
5. Log every model and tool call with agent and user identity.
6. Define kill, quarantine and rollback procedures and test them.
7. Feed agent evidence into your continuous monitoring, not only the initial ATO package.

Our pillar on [mapping AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how to reuse the same evidence across FedRAMP, SOC 2 and ISO 42001.

## How Opencontroller does this

Opencontroller discovers agents, models and API keys across AWS, Azure and Google Cloud, keeps a governed registry of agents, versions, tools and MCP servers, and detects risky permissions and configurations with Posture. Delegated on-behalf-of identity and agent workload identity federation tie calls to identities, and rogue-agent handling supports quarantine and kill. It can run on customer-controlled infrastructure, which keeps it inside the agency's own boundary. Policy as code ships with built-in mappings to NIST, among others.

## Frequently asked questions

**Does an AI agent need its own FedRAMP authorisation?**
Usually not. FedRAMP authorises cloud service offerings. An agent built on an authorised service is covered by the agency's own ATO for the system that uses it, which must address the agent's added risks.

**Is a FedRAMP-authorised model API enough to deploy agents?**
No. It covers the provider's boundary. The agency must still govern the agent's identity, tool access, data flows and logs.

**What does FedRAMP 20x change for AI agents?**
It moves evidence toward automated, machine-readable indicators, which fits fast-changing agent estates better than periodic narrative documentation.

To see continuous agent evidence in your environment, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

FedRAMP (run by GSA) authorises cloud service offerings, not individual AI agents; agencies still need their own ATO covering agents built on authorised services. FedRAMP's AI Prioritization Initiative began in August 2025, was scoped to conversational AI engines for routine use by federal workers, required five criteria including SSO/SCIM/RBAC, demand from five CFO Act agencies or a CIO Council recommendation, GSA MAS availability and meeting FedRAMP 20x within two months, and completed in April 2026. FedRAMP 20x moves authorisation toward automated, machine-readable evidence via Key Security Indicators. Opencontroller provides evidence for agent inventory, identity, posture, logging and kill/quarantine and can run on customer-controlled infrastructure; it does not claim FedRAMP authorisation.

## Related

- [What does OMB M-25-21 require for AI agents in US federal agencies?](https://opencontroller.lyzr.com/blog/omb-m-25-21-and-federal-ai-agents.md)
- [What are NIST control overlays for AI systems and agents?](https://opencontroller.lyzr.com/blog/nist-control-overlays-for-ai-agents.md)
- [What SOC 2 evidence do auditors want for AI agents?](https://opencontroller.lyzr.com/blog/soc-2-evidence-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
