grasp_agents 0.5.24

Grasp Agents Library

Grasp Agents

Overview

Grasp Agents is a modular Python framework for building agentic AI pipelines and applications. It is meant to be minimalistic but functional, allowing for rapid experimentation while keeping full and granular low-level control over prompting, LLM handling, and inter-agent communication by avoiding excessive higher-level abstractions.

Features

• Clean formulation of agents as generic entities over I/O schemas and shared context.
• Transparent implementation of common agentic patterns:
  • Single-agent loops
  • Workflows (static communication topology), including loops
  • Agents-as-tools for task delegation
  • Freeform A2A communication via the in-process actor model
• Built-in parallel processing with flexible retries and rate limiting.
• Support for all popular API providers via LiteLLM.
• Granular event streaming with separate events for standard outputs, thinking, and tool calls.
• Callbacks via decorators or subclassing for straightforward customisation of agentic loops and context management.

Project Structure

• processors/, llm_agent.py: Core processor and agent class implementations.
• packet.py, packet_pool.py, runner.py: Communication management.
• llm_policy_executor.py: LLM actions and tool call loops.
• prompt_builder.py: Tools for constructing prompts.
• workflow/: Modules for defining and managing static agent workflows.
• llm.py, cloud_llm.py: LLM integration and base LLM functionalities.
• openai/: Modules specific to OpenAI API integration.
• litellm/: Modules specific to LiteLLM integration.
• memory.py, llm_agent_memory.py: Basic agent memory management.
• run_context.py: Shared context management for agent runs.
• usage_tracker.py: Tracking of API usage and costs.
• rate_limiting/: Basic rate limiting tools.

Quickstart & Installation Variants (UV Package manager)

Note: You can check this sample project code in the src/grasp_agents/examples/demo/uv folder. Feel free to copy and paste the code from there to a separate project. There are also examples for other package managers.

1. Prerequisites

Install the UV Package Manager:

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Create Project & Install Dependencies

mkdir my-test-uv-app
cd my-test-uv-app
uv init .

Create and activate a virtual environment:

uv venv
source .venv/bin/activate

Add and sync dependencies:

uv add grasp_agents
uv sync

3. Example Usage

Ensure you have a .env file with your OpenAI and Google AI Studio API keys set

OPENAI_API_KEY=your_openai_api_key
GEMINI_API_KEY=your_gemini_api_key

Create a script, e.g., problem_recommender.py:

import asyncio
from typing import Any

from dotenv import load_dotenv
from pydantic import BaseModel, Field

from grasp_agents import LLMAgent, BaseTool, RunContext, Printer
from grasp_agents.litellm import LiteLLM, LiteLLMSettings


load_dotenv()

sys_prompt_react = """
Your task is to suggest an exciting stats problem to the student. 
You should first ask the student about their education, interests, and preferences, then suggest a problem tailored specifically to them. 

# Instructions
* Use the provided tool to ask questions.
* Ask questions one by one.
* Provide your thinking before asking a question and after receiving a reply.
* Do not include your exact question as part of your thinking.
* The problem must have all the necessary data.
* Use the final answer tool to provide the problem.
"""

# Tool input must be a Pydantic model to infer the JSON schema used by the LLM APIs
class TeacherQuestion(BaseModel):
    question: str


StudentReply = str


ask_student_tool_description = """
"Ask the student a question and get their reply."

Args:
    question: str
        The question to ask the student.
Returns:
    reply: str
        The student's reply to the question.
"""


class AskStudentTool(BaseTool[TeacherQuestion, StudentReply, None]):
    name: str = "ask_student"
    description: str = ask_student_tool_description

    async def run(self, inp: TeacherQuestion, **kwargs: Any) -> StudentReply:
        return input(inp.question)


class Problem(BaseModel):
    problem: str


teacher = LLMAgent[None, Problem, None](
    name="teacher",
    llm=LiteLLM(model_name="gpt-4.1"),
    tools=[AskStudentTool()],
    react_mode=True,
    final_answer_as_tool_call=True,
    sys_prompt=sys_prompt_react,
)

async def main():
    ctx = RunContext[None](printer=Printer())
    out = await teacher.run("start", ctx=ctx)
    print(out.payloads[0])
    print(ctx.usage_tracker.total_usage)


asyncio.run(main())

Run your script:

uv run problem_recommender.py

You can find more examples in src/grasp_agents/examples/notebooks/agents_demo.ipynb.