Dana (Domain-Aware NeuroSymbolic Architecture)¶
Vision¶
Dana is a universal program format and execution runtime that enables intelligent agents — human or machine — to reason, act, and collaborate through structured, interpretable programs.
It serves as the missing link between natural language objectives and tool-assisted, stateful action. Dana programs are concise, auditable, explainable, and can be authored by LLMs, domain experts, or both.
Motivation & Problem¶
Modern AI systems struggle with:
- Prompt chains are fragile — hard to debug, hard to maintain
- Plans are opaque — impossible to inspect or explain mid-flight
- Tool use is scattered — logic is buried in code, not declarative programs
- State is implicit — no shared memory model or traceable updates
Symbolic systems offer structure but lack adaptability. LLMs offer creativity but lack transparency. Dana bridges the two.
✅ Solution¶
Dana introduces a lightweight domain-aware program language and runtime. It allows:
- Programs as first-class reasoning artifacts
- Shared state containers (
agent,world,temp,execution) - Reusable logic units via a structured Knowledge Base (KB)
- Declarative goals, imperative execution
- Bidirectional mapping to/from natural language
Dana can:
- Be generated by a planning agent (like GMA)
- Be executed line-by-line by a runtime
- Interact with tools, LLMs, and memory
- Be stored, versioned, tested, and explained
Architecture Overview¶
Emitters and Interpreters of Dana¶
| Actor | Type | Role(s) in Dana | Description |
|---|---|---|---|
| User (Human) | Person | 🖋 Emitter | Writes Dana directly to define goals, logic, or KB entries |
| GMA | Agent | 🖋 Emitter | General planner that emits Dana plans from objectives |
| DXA | Domain Agent | 🖋 Emitter | Emits specialized domain logic/workflows, often tied to KB content |
| KB Maintainer | Person or Agent | 🖋 Emitter | Curates reusable Dana programs as structured knowledge |
| Tool Resource | System Component | ✅ Interpreter | Executes atomic tool-backed actions referenced in Dana |
| Local Runtime | System Component | ✅ Interpreter | Executes Dana deterministically except for reason(...) |
| Dana_LLM | LLM Wrapper Module | 🖋 Emitter + ✅ Interpreter | Emits code and executes reasoning operations |
| AgentRuntime | System Component | 🔁 Coordinator | Orchestrates execution and manages delegation across all actors |
State Model¶
Dana programs operate over a shared RuntimeContext, which is composed of four memory scopes (state containers):
| Scope | Description |
|---|---|
local: |
Local to the current agent/resource/tool/function (default scope) |
private: |
Private to the agent, resource, or tool itself |
public: |
Openly accessible world state (time, weather, etc.) |
system: |
System-related mechanical state with controlled access |
Note: Only these four scopes are valid in the Dana language and enforced by the parser. Any references to other scopes (such as
agent:,world:,temp:,stmem:,ltmem:,execution:, or custom scopes) are not supported in the current grammar and will result in a parse error.
Security Design¶
**The dana.runtime