Agent Habitat Architecture

A proposed security architecture for enterprise AI agents.

Explore principles for environments where agents run code and use tools, plus maturity tracks for evaluating your security controls.

Cover of the Agent Habitat Architecture white paper from Replit

Get the whitepaper

See the full picture.

Core principles, a proposed architecture, and a maturity model.

Loading the form…

What's inside

  1. A blueprint for agents you can trust with real work.

    Security boundaries designed to hold even when an agent is manipulated or actively looking for gaps.

  2. Controls that go beyond “tell the agent not to.”

    Permissions that only shrink, credentials agents never see, and consequential actions that require human approval.

  3. A practical way to separate security claims from proof.

    A maturity scale with testable evidence for assessing your environment or evaluating vendors.

  4. Real-world failures turned into actionable lessons.

    How agent systems break, what would have stopped them, and where to strengthen your defenses first.

A closer look

Learn about practical frameworks

Discover how secure agent habitats should be built. The white paper complements industry efforts like the NVIDIA Open Agent Safety Platform and the Blueprint Alliance reference architecture, and goes into more detail for practical application.

Get the whitepaper
The paper's proposed architecture diagram: a trusted zone with the user's input and human approval checkpoints; an identity provider; a harness that launches sandboxes in an untrusted execution zone; and a trusted control plane where a token service, vault, credential broker and policy service decide which outbound requests reach the LLM, internet services and internal services.