Rasa Mantle
A conversational agent engine introduced by Rasa that declares skills, tools, and memory in natural language, supporting tool calling, memory, and abstraction over the LLM layer.
Timeline (last 90 days)
Rasa Mantle is Rasa's newly launched agent engine, and its three core capabilities include tool calling, memory, and the LLM layer.
In the NoteVs project, the Mantle version needs about 210 lines of configuration, whereas the classic CALM framework implementation of the same functionality requires 5 files and about 590 lines of YAML.
In Mantle, tools are async Python functions with decorators, and type hints and descriptions are the only basis for the LLM to understand a tool's purpose.
The author has tools return a ToolResult containing the error code instead of throwing an exception when an HTTP request fails, so that the LLM can give a readable error reply and users do not see stack traces.
The way to fix multi-window project confusion is to have the helper function explicitly pass the project folder path as a parameter on every call, with that path coming from the environment variable each window sets for its agent.
Mantle skills list the available tools via front matter and can have constraints enforced by the framework via tool_constraints, for example delete, remind, and export operations must wait for user confirmation.
Mantle's built-in session-start skill greets the user and ends the turn, causing the first message of every conversation to be swallowed; the author solved the problem by overriding this behavior with a skill that does nothing.
Mantle's memory entries must first be declared in memory.yml, and the LLM can only write entries marked as llm_settable, so it cannot forge critical values such as verification results.
All Facts (8)
Rasa Mantle is Rasa's newly launched agent engine, and its three core capabilities include tool calling, memory, and the LLM layer.
50%UnverifiedIn the NoteVs project, the Mantle version needs about 210 lines of configuration, whereas the classic CALM framework implementation of the same functionality requires 5 files and about 590 lines of YAML.
50%UnverifiedIn Mantle, tools are async Python functions with decorators, and type hints and descriptions are the only basis for the LLM to understand a tool's purpose.
50%UnverifiedThe author has tools return a ToolResult containing the error code instead of throwing an exception when an HTTP request fails, so that the LLM can give a readable error reply and users do not see stack traces.
50%UnverifiedThe way to fix multi-window project confusion is to have the helper function explicitly pass the project folder path as a parameter on every call, with that path coming from the environment variable each window sets for its agent.
50%UnverifiedMantle skills list the available tools via front matter and can have constraints enforced by the framework via tool_constraints, for example delete, remind, and export operations must wait for user confirmation.
50%UnverifiedMantle's built-in session-start skill greets the user and ends the turn, causing the first message of every conversation to be swallowed; the author solved the problem by overriding this behavior with a skill that does nothing.
50%UnverifiedMantle's memory entries must first be declared in memory.yml, and the LLM can only write entries marked as llm_settable, so it cannot forge critical values such as verification results.
50%