Design Principles

Isolation

• Donews don’t need to know about each other’s inner workings.
• They don’t taint each other’s state or prompts.
• They can be run in parallel and information is collected and combined once available. (tbd)
• Donews can easily be tested on their own.
• This makes debugging and testing much easier and protects your sanity.

Natural Design Language

• We deliberately chose terms that separate known coding concepts and “agentic”(you will hear us saying this only once. and it is here) concepts.
• You define a donew task by providing the following concepts:
  • realm the provision pool that the overseeing donew can use to accomplish the task (e.g. “web browsing”, “reading documentation” etc)
  • envision a defined structured outcome. (e.g. a pydantic model or a looser structured outcome)
  • enact starts the task (e.g. “find the best pizza in SF”).
• realm and envision are optional, enact is required.

Batteries Included

• We provide battle tested pre-built realms for common tasks.
• We partner with the best in class libraries such as mcp.run to provide powerful integrations.
• We have a built in LLM tracing system that allows you to understand the inner workings of the provision pool.

Extensible and Composable

• Donew is designed to be extendable.
• You can define your own realms, envisionments and enactments.
• You can use the donew SDK to build your own integrations.
• Putting donew tasks into the realm of another donew task allows you to compose complex workflows.

Reproducible

• We are working hard to build a chrystalization module that will allow you to take any task and break it down into a set of discrete steps.
• Meaning once the donews have successfully completed a task, the steps can be used to reproduce the same outcome with varying inputs.
• In the background all of our donews write code to accomplish any task.
• This allows us to bridge the gap between autonomous problem solving and creating reusable code.