Build agentic systems in a layered, expressive, and secure architecture
AI proposes. Your code decides.
Propose, Rehearse, Commit.
The optimal architecture for agents
Agents follow the user's direction to propose actions. Deterministic code validates, rehearses, and commits the proposed changes.
Go to libraries and examples01 / 03
People set the direction.
This is the interface where people work: writing, designing, organizing, or solving a problem. The application records their actions and choices as structured data.
The user stays in charge of what they want to achieve.
02 / 03
A copilot makes proposals.
The agent observes the user’s actions, choices and recent context, then proposes actions that support the user’s direction and goals. It turns an understanding of the task into typed proposals over the operations the application exposes.
The model proposes; deterministic software decides which changes can take effect.
03 / 03
Deterministic software does the work.
This layer uses classic software design patterns and engineering practices. It can be an existing codebase, with a selective Programming Interface that exposes deterministic operations to the agentic layer above.
The runtime validates the proposed work, rehearses changes in an isolated copy, then commits when all the required checks pass.
THE COMPLETE PICTURE
Direction, proposals, and controlled effects.
The three planes assemble into the Sergent architecture. Each layer has a distinct responsibility, and together they form the Sergent design pattern.
For more information, start with the Sergent Specification.
If you would like to see the Sergent architecture and design pattern in action, explore the libraries and examples below.
The specification, libraries, and examples live on GitHub. Choose your programming language, then use the library’s documentation and its companion examples to get started.
| Language | Library | Examples |
|---|---|---|
| Python | sergent-py | sergent-py-examples |
| Go | sergent-go | sergent-go-examples |
| Rust | sergent-rs | sergent-rs-examples |
| TypeScript | sergent-ts | sergent-ts-examples |
Human vs. AI or AI vs. AI for both fun and AI intelligence evaluation.
A human drives content creation with a team of AI assistants filling in the gaps.
Break a complex subject down and then iterate on all the elements to create a full answer.
AI writes code in a secure sandbox with a built-in adversarial review process.
Add agentic decision-making to system programs where reliability and portability matter.
Run an agentic application in the browser without any external model providers.
Coding agents can read this page as plain text , then follow the same repository links. Installation and API details belong to each library’s documentation.