Introducing Guard for Agents

With Command, we delivered a single control center for developers to see, test and manage how ZeroDrift applies compliance policy across AI-generated communication.

Today, we're extending that control directly into the agent workflow with Guard for Agents, making it easier to connect the policies an agent needs to follow and put that enforcement directly into the agent workflow.

Developers can connect policy sources including Notion, Linear and Google Drive through MCP. Guard extracts the policies teams have already written and turns them into enforceable rules. The Validation API gives developers a direct way to run those rules against agent-generated communication, while the OpenAPI specification makes it possible to generate typed clients in languages including Python, TypeScript, Go and Java.

Every enforcement check runs on Anchor 3.0-Preview, ZeroDrift's purpose-built model for detecting policy violations and generating compliant rewrites rather than a general-purpose model prompted with your rulebook on every request. Guard for Agents is the first place developers can put the model to work.

As agents move into more workflows, policy must move with them

Enterprises are pushing AI agents into more functions and more customer-facing workflows. For regulated industries in particular, every new place an agent communicates introduces another place where regulatory requirements and company policy must be enforced.

Much of the enterprise AI governance conversation has focused on access: systems an agent can reach, tools it can call, and actions it's allowed to take. Those controls matter, but the thing a customer actually experiences is the message.

The challenge is not necessarily defining those policies. A company may already know exactly what an agent should and should not say, but those requirements do not automatically follow an agent into every new application or workflow. Developers need a consistent way to connect the two without treating compliance as a separate implementation problem every time an agent is deployed somewhere new.

Put your policies in the path of the agent

Take a wealth management firm building an AI assistant that communicates with customers. The policies governing what that assistant can say may already live in Google Drive, while the assistant itself operates somewhere else entirely.

With Guard for Agents, developers can connect those policy documents to ZeroDrift through MCP, turn the policies inside them into enforceable rules and use the Validation API to apply those rules directly in the assistant's response flow.

If the assistant generates a message that violates one of those rules, Guard catches it before delivery and returns the appropriate action. That can mean providing a compliant replacement for AI-generated content, blocking the message or escalating it for review, depending on the policy that applies.

What ships today

Guard for Agents ships today with three things developers can use immediately:

• A production Validation API
• A full OpenAPI specification
• MCP connectors for turning existing policy documents into enforceable rules

An API built to run in production, today

The whole integration is a single POST to our validation endpoint — no infrastructure to stand up, no batch job to schedule. Call it inline in the same request path as your agent's response, and it's fast enough to sit there.

Send the text your agent wants to output. The response identifies the rule that was violated, its severity, and — for AI-generated content — a compliant replacement your application can send instead. If something's too serious to fix automatically, the response tells you to block it or escalate it instead. Humans are never auto-rewritten this way; auto-fix is for machine-generated content only.

Generate a client in whatever language you use

The complete Validation API is published as an OpenAPI specification. Generate typed clients in Python, TypeScript, Go or Java, create a Postman collection or hand the spec directly to your coding agent. There's no SDK release cycle to wait on. The specification is the source of truth and ships alongside the API.

MCP connectors that ingest your policy docs and turn them into rules

Guard for Agents ships with MCP connectors for Notion, Linear and Google Drive, with more on the way. Connect one and Guard ingests your policy documents, extracts the rules and runs them live against agent communications.

The compliance handbook sitting in a shared drive stops being something people are supposed to remember and becomes something enforced on every message your agent sends. Your policy stops being documentation. It becomes runtime infrastructure.

Autonomous doesn't mean unaccountable

ZeroDrift logs every enforcement decision, including the original message, the rule that was triggered when a message is flagged or blocked, and the action Guard took and the output that ultimately passed through.

And every rule you instate stays under your control, meaning you can turn them on or off at runtime, configure them per pipe or per agent, and see the impact immediately.

The enforcement is autonomous, but the reasoning behind it isn’t hidden.

Guard for Agents is live today. The docs are up now, with examples you can run in minutes.

‍