decionis/agent-safe-pipeline is an open-source reference architecture for building AI agents that can propose actions but are prevented from executing them directly, enforcing a safety-by-design principle. It provides a structured framework to separate an agent’s decision-making logic from its ability to perform potentially unsafe operations, requiring human or system-level approval for execution. This pattern is designed to mitigate risks associated with autonomous AI systems by ensuring that all proposed actions undergo a validation step.
What it is
The decionis/agent-safe-pipeline is a blueprint for developers building AI agents, particularly those that interact with external systems or APIs. The core concept is to architecturally enforce a separation of concerns: the AI agent operates within a constrained environment where it can analyze situations and suggest actions, but the actual execution of those actions is handled by a separate, secure component. This prevents the AI from taking unintended or harmful actions autonomously.
Implemented in TypeScript, the project offers a practical example of how to structure code to achieve this safety goal. By providing a working example, it demonstrates how to implement the pipeline, including how an agent’s proposals are captured, reviewed, and then conditionally executed by a trusted system. This approach is critical for applications where an AI’s error could have significant consequences.
Key facts
| Attribute | Details |
| :— | :— |
| Released | v0.1.0 (August 15, 2026) |
| Developer | decionis |
| License | Apache-2.0 |
| Primary Language | TypeScript |
| Availability | GitHub repository: `decionis/agent-safe-pipeline` |
How it compares
The decionis/agent-safe-pipeline is a specific implementation of a safety pattern for AI agents, distinct from general-purpose AI frameworks like LangChain or LlamaIndex. While those frameworks provide tools for building agent capabilities, this project focuses exclusively on the architectural pattern of separating action proposal from authorization. It is more comparable to other safety-focused AI projects but is defined by its specific, lightweight reference implementation in TypeScript.
FAQ
What problem does this architecture solve?
It addresses the security risk of AI agents autonomously executing harmful or incorrect actions. By forcing a separation between an agent’s “thinking” (proposing actions) and “acting” (executing them), it ensures a human or a secure system component must approve every action, significantly reducing the potential for unintended consequences.
How do I get started with the project?
The project can be cloned from its GitHub repository. The recommended installation process involves using `git clone` to download the code, followed by using `pnpm` to install dependencies and run a basic demonstration to see the safety pipeline in action.
Is this project production-ready?
As a reference architecture with an initial v0.1.0 release, it is intended primarily as an educational and foundational blueprint. Developers would use this code as a starting point to understand and implement the safety pattern, adapting and hardening it for their specific production needs.
This definition is written from primary sources — the model card, repository, and official documentation — not paraphrased from other summaries. It is updated as the term’s details change.