AmritaSense Overview

We recap the pieces of AmritaSense that AmritaCore builds on. Full documentation lives at sense.amritabot.com.

The Idea

AmritaSense compiles workflows into a linear instruction sequence executed by a lightweight VM — much like a CPU runs machine code. Nodes are chained with >>; control flow is native instructions (IF, WHILE, GOTO, CALL, TRY, NOP).

from amrita_sense import Node, WorkflowInterpreter


@Node()
async def step_one() -> None:
    print("[1] load state")


@Node()
async def step_two() -> None:
    print("[2] process")


composition = step_one >> step_two
interpreter = WorkflowInterpreter(composition.render())
await interpreter.run()

What Core Uses from Sense

Sense primitive	Where Core uses it
@Node	Every component (components/llm.py, process.py, react.py)
WorkflowInterpreter	ChatObject._interpreter runs the conversation pipeline
Dependency injection (type-matched)	Workflow nodes receive AgentLoopState, AbilityState, ...
SuspendObjectStream	ChatObject.io_stream — bidirectional streaming
Matcher events	The pipeline/step hook system (see Events)
NATIVE instructions	The built-in step loop (NATIVE_DO/NATIVE_WHILE)

The VM

• Program counter (PointerVector) + call stack drive execution
• Nodes resolve dependencies before running (DI)
• Every node boundary catches exceptions
• run_step_by() yields each step for debugging

See Workflow Engine for how ChatObject composes its pipeline, and sense.amritabot.com for the engine reference.