Skip to main content

​
The LIATE Specification

OpenLiate decomposes agent design into five isolated, modular layers. In agent.json, each layer corresponds to a top-level key:
agent.json

[L] Large Models Layer

Provider abstraction supporting Indian sovereign models (Sarvam AI), Ollama, Anthropic, and OpenAI.

[I] Integration & Memory Layer

Multi-tier session persistence: In-Memory, File-backed JSON, and Vector DB integration.

[A] Agent Harness & Loop Layer

Autonomous ReAct execution loop (Thought →\rightarrow Action →\rightarrow Observation →\rightarrow Answer) with loop guardrails.

[T] Tools & MCP v2 Layer

Built-in identity tools (Aadhaar & PAN) and Model Context Protocol (MCP) tool server dispatching.

[E] Environment & Sandbox Layer

Secret masking (sk-...), safe $VAR expansion, domain whitelisting, and execution guardrails.