A lightweight LLM Agentic framework with a four-layer abstraction (Agent/Prompt/Dialog/Tactic) for rapid prototyping and multi-agent orchestration, featuring built-in Proxy system, sandbox execution, and structured logging.
LLLM (Low-Level Language Model) is a lightweight Python framework designed for building advanced LLM Agentic systems, with explicit support for program synthesis and neural-symbolic research. Developed by the Productive-Superintelligence organization, latest version v0.1.1, licensed under Apache-2.0.
Core Abstractions#
The framework is built on a four-layer abstraction:
- Agent (caller): Encapsulates system prompts, base models, and tool call loops
- Prompt (function): Manages templates, parsers, and processors
- Dialog (internal state): Maintains independent conversation histories with fork support
- Tactic (program): Orchestrates Agent-Prompt collaboration logic
Rapid Prototyping#
Tactic.quick() enables one-line conversations with zero configuration:
from lllm import Tactic
response = Tactic.quick("What is the capital of France?")
print(response.content)
Multi-Model & Tools#
Model calls are unified through LiteLLM, supporting OpenAI GPT-4o, Anthropic Claude, and other providers. The @tool decorator enables function calling, with built-in run_python and query_api_doc tools.
Proxy System & Sandbox#
Built-in agent modules (financial data, search, etc.) with auto-registration, injecting tools and sandbox execution environments (interpreter/Jupyter) where variables persist across calls.
Structured Output & Logging#
Pydantic model-driven Prompt formatting and output parsing. LogStore supports LocalFile/SQLite/NoOp backends with tag indexing, cost aggregation, and export.
Configuration & Package Management#
lllm.toml project config and YAML Agent config enable declarative multi-agent orchestration with inheritance and deep merging. lllm pkg install enables shareable, versioned Agent infrastructure.
Batch & Concurrency#
bcall()/acall()/ccall() support parallel/batch execution. load_runtime()/get_runtime() support named runtime parallel experiments.
Experimental Features#
Playwright-based Computer Use Agent (CUA), OpenAI Responses API routing, and a Skills system.
Use Cases#
Rapid prototyping, multi-agent collaboration systems, proxy-driven data analysis, code review service packaging (built-in FastAPI example), session auditing & cost tracking, program synthesis & neural-symbolic research.
Installation#
pip install lllm-core