decionis/agent-safe-pipeline

What it solves

AgentSafe acts as an authority boundary to prevent autonomous agents or applications from performing unauthorized or dangerous actions. It addresses the risk where an agent with valid credentials might be compromised or hallucinate, performing actions that exceed its intended scope or authority.

How it works

AgentSafe sits as a proxy between your agent and the target service. When the agent attempts an action, AgentSafe intercepts the HTTP request, normalizes it into an "intent," and queries the Decionis control plane for a decision. Based on organizational policy, the control plane returns an ALLOW, BLOCK, or ESCALATE (which triggers a human verification ceremony). If authorized, AgentSafe forwards the request exactly once using a single-use execution grant and records the entire transaction as a signed "Decision Dossier."

Who it’s for

This tool is designed for developers and security engineers who need to govern autonomous agents, LLM-based applications, or automated infrastructure tools by enforcing strict, action-level authorization that goes beyond simple identity-based access control.

Highlights

• Action-level authorization: Decisions are bound to the specific action (e.g., the exact payment amount or resource scale) rather than just the agent's identity.
• Human-in-the-loop: Supports ESCALATE workflows where consequential actions require manual approval via a verified human presence layer.
• Shadow mode: Allows you to observe what the system would block or allow without actually interrupting your traffic, facilitating safe adoption.
• Adversarial testing: Includes built-in test suites to verify that the boundary correctly handles forged headers, replayed grants, and unauthorized parameter changes.
• Flexible deployment: Can be deployed as a proxy, a sidecar in Kubernetes, or via transparent network interception.