agentloop

A small, reusable Python agentic framework for building LLM agents with tools, memory, and an observe–act loop. Other developers can extend it by registering tools, swapping LLM providers, and hooking lifecycle events—without pulling in a heavy dependency stack.

Features

• Agent loop — model → tool calls → tool results → repeat until done or max_iterations
• Tool registry — decorate functions; JSON schema inferred from type hints
• Pluggable LLM providers — OpenAI-compatible API built-in; implement one method for custom backends
• Memory — conversation history with optional system prompt
• Events — on_tool_call, on_iteration, on_run_end, generic subscribe() hooks
• Mock provider — run examples and tests without an API key

Install

From PyPI (after you publish — see PUBLISHING.md):

pip install agentloop-framework
pip install "agentloop-framework[openai]"   # OpenAI-compatible provider

From source (development):

cd agentloop
pip install -e .

# For OpenAI / Azure / Ollama (OpenAI-compatible HTTP API):
pip install -e ".[openai]"

# Development:
pip install -e ".[dev]"

The PyPI name is agentloop-framework because agentloop is already taken. Imports remain from agentloop import ....

Quick start (no API key)

import asyncio
from agentloop import Agent
from agentloop.providers import MockProvider
from agentloop.tools import ToolRegistry

tools = ToolRegistry()

@tools.register(description="Echo text back.")
def echo(text: str) -> str:
    return text

agent = Agent.create(
    provider=MockProvider(MockProvider.tool_then_answer("echo", {"text": "hi"}, final="Done.")),
    system="You are helpful.",
    tools=tools,
)

asyncio.run(agent.run("Say hi"))

Run the bundled example:

python examples/basic_mock.py

Example agents

Chatbot (multi-turn)

Interactive CLI chat with persistent memory and tools (time, calculator):

python examples/chatbot.py --mock          # offline, no API key
python examples/chatbot.py --live          # needs OPENAI_API_KEY

Voice agent

Speak to the agent and hear replies (or use --text to type):

python examples/voice_agent.py --text --mock   # easiest start
pip install -e ".[voice]"
python examples/voice_agent.py --mock          # microphone + speaker

See examples/README.md for full details.

Cloud Run + Pinecone RAG + React UI

Production-style deployment with vector search:

# See deploy/cloud-run/README.md for full steps
docker build -f deploy/cloud-run/Dockerfile -t agentloop-chat .

Stack: React chat UI → FastAPI /api/chat → agentloop → Pinecone on Google Cloud Run.

Live LLM (OpenAI-compatible)

import asyncio
from agentloop import Agent
from agentloop.providers import OpenAICompatibleProvider
from agentloop.tools import ToolRegistry

tools = ToolRegistry()

@tools.register(description="Add two integers.")
def add(a: int, b: int) -> int:
    return a + b

agent = Agent.create(
    provider=OpenAICompatibleProvider(model="gpt-4o-mini"),
    system="Use tools when math is needed.",
    tools=tools,
)

asyncio.run(agent.run("What is 12 + 30?"))

Environment variables:

Architecture

flowchart LR
  User[User task] --> Loop[AgentLoop]
  Loop --> LLM[LLMProvider.chat]
  LLM -->|tool_calls| Tools[ToolRegistry.execute]
  Tools --> Memory[ConversationMemory]
  Memory --> Loop
  LLM -->|final text| Result[RunResult]

Extension points

Project layout

agentloop/
  agent.py          # Agent.create() builder
  loop.py           # Core observe–act loop
  tools.py          # ToolRegistry
  memory.py         # ConversationMemory
  events.py         # AgentEvents
  types.py          # Message, ToolCall, RunResult, ...
  providers/
    base.py         # LLMProvider protocol
    openai_compatible.py
    mock.py
examples/
  chatbot.py        # Multi-turn interactive chat
  voice_agent.py    # Speech input / spoken output
  helpers.py        # Shared tools + agent factory
  README.md
tests/

Tests

pytest

Design choices

• Minimal core — zero required runtime dependencies; httpx only for the optional OpenAI provider.
• Async-first — asyncio throughout; sync wrappers can wrap asyncio.run in your app.
• Provider-agnostic — not tied to Cursor or any single vendor; bring your own model endpoint.

License

MIT