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 import DotDict
from appinfra.log import create_lg

from llm_gent import Agent, LLMTrait
from llm_gent.core.agent.types import ExecutionResult


# The public Agent class is abstract — a real application defines a small
# concrete subclass. Trivial stubs suffice when the workflow only uses
# LLMTrait.complete() directly.
class HelloAgent(Agent):
    def start(self) -> None:
        self._start_traits()

    def stop(self) -> None:
        self._stop_traits()

    def run_once(self) -> ExecutionResult:
        return ExecutionResult(success=True, content="")

    def ask(self, question: str) -> str:
        return ""

    def record_feedback(self, message: str) -> None:
        pass

    def get_recent_results(self, limit: int = 10) -> list[ExecutionResult]:
        return []


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

# Agent reads config.identity.name internally.
config = {"identity": {"name": "hello-agent"}}

llm_config = DotDict(
    {
        "default": "local",
        "backends": {
            "local": {
                "type": "openai_compatible",
                "base_url": "http://localhost:8000/v1",
                "model": "default",
            }
        },
    }
)

agent = HelloAgent(lg, config)
agent.add_trait(LLMTrait(agent, llm_config))
agent.start()

# LLMTrait.complete() accepts OpenAI-style message dicts.
llm = agent.require_trait(LLMTrait)
result = llm.complete(
    [
        {"role": "system", "content": "You are a concise assistant."},
        {"role": "user", "content": "Say hello."},
    ]
)
print(result.content)

agent.stop()

A runnable version of this example lives at llm_gent/examples/quickstart.py. Set LLM_GENT_SMOKE=1 to run it against a stub LLM router (used by CI's wheel-smoke job).

Core Concepts

Agents

An Agent is a container for traits with lifecycle management. Agents have an identity (domain/workspace/name) and can be started, stopped, and 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
llm-gent serve

# Or with specific config
llm-gent -c etc/llm-gent.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.