llm-gent

Agent framework with trait-based architecture and learning capabilities.

Overview

llm-gent provides a composable framework for building LLM-powered agents. Agents are composed of traits that provide specific capabilities (LLM access, storage, learning, etc.) and can be run standalone or as services via the included HTTP runtime.

Key features:

• Trait-based composition - Mix and match capabilities via traits (LLM, Storage, Rating, Learn)
• Multi-backend LLM support - OpenAI-compatible, Anthropic, and custom backends via llm-infer
• Built-in learning - Collect training data (SFT/DPO) and fine-tune via llm-kelt
• Structured output - Pydantic schema validation with automatic JSON cleanup for small models
• Production ready - HTTP server, PostgreSQL storage, schema migrations

Installation

pip install llm-gent

For HTTP server support:

pip install llm-gent[http]

Supported Python versions

CI tests against Python 3.11, 3.12, 3.13, and 3.14 on every push. requires-python = ">=3.11".

Quick Start

from appinfra.log import create_lg

from llm_gent import AgentFactory, LLMTrait

lg = create_lg("my-agent", "info")

agent = AgentFactory(lg).from_config(
    {
        "identity": {"name": "my-agent"},
        "llm": {
            "default": "local",
            "backends": {
                "local": {
                    "type": "openai_compatible",
                    "base_url": "http://localhost:8000/v1",
                    "model": "default",
                }
            },
        },
        "directive": "You are a helpful assistant.",
        "traits": {"required": ["llm", "directive"]},
    }
)

agent.start()
llm = agent.require_trait(LLMTrait)
result = llm.complete([{"role": "user", "content": "Hello!"}])
print(result.content)
agent.stop()

A runnable version of this example lives at llm_gent/examples/quickstart.py. Pass --smoke to run it against a stub LLM router (no real backend contacted).

Core Concepts

Agents

An Agent is a container for traits with lifecycle management. Agents have an identity (name plus optional context_key) and can be started, stopped, and — via RunnableAgent — run in cycles.

Traits

Traits provide specific capabilities to agents:

Tools

Built-in tools for agentic workflows:

• ShellTool - Execute shell commands
• FileReadTool / FileWriteTool - File operations
• HTTPFetchTool - HTTP requests
• RecallTool / RememberTool - Memory operations

Running as a Service

# Start agent server (uses packaged default config)
llm-gent serve

# With custom config
llm-gent --config path/to/config.yaml serve

Related Projects

• llm-infer - LLM inference server and client
• llm-kelt - Training infrastructure (SFT/DPO)
• appinfra - Application infrastructure utilities

License

Apache License 2.0 - see LICENSE for details.

Maintained by LLM Works LLC and contributors.