pico-agent

DEPRECATED. This package is no longer maintained and receives no further releases. Version 0.2.1 is the final release; it fixes compatibility with pico-ioc >= 2.4.0 so existing installs keep working. The rest of the pico ecosystem is unaffected and actively maintained.

Pico-Agent

Pico-Agent is a lightweight, protocol-based framework for building modular AI agents using dependency injection.

It eliminates the boilerplate of manually managing LLM chains, prompts, and tools. Instead, you declare what you want the agent to do using Python Protocols, and the framework handles how to execute it via Pico-IoC.

Protocol-First Design True Dependency Injection Smart Model Routing Multi-Agent Orchestration LLM Agnostic (LangChain based)

Why pico-agent?

Most agent frameworks require you to instantiate classes, manually bind tools, and hardcode model names. Pico-Agent moves this complexity into the IoC container, promoting clean architecture and testability.

Core Features

• Declarative Agents: Use @agent on standard typing.Protocol.
• Capability Routing: Request FAST, SMART, CODING capabilities instead of specific model names.
• Auto-Discovery: Leverages pico-ioc 2.2+ custom scanners to find agents automatically.
• Type-Safe: Full support for Pydantic input/output schemas.
• ReAct Loops: Built-in support for reasoning loops and tool usage.

Installation

pip install pico-agent

Install with specific provider support:

pip install "pico-agent[openai,google]"

Quick Example

1. Define the Agent Protocol

The method signature defines the tool input. The @agent decorator configures everything else.

from typing import Protocol
from pico_agent import agent, AgentCapability, AgentType

@agent(
    name="translator",
    capability=AgentCapability.FAST,  # Uses a faster/cheaper model
    system_prompt="You are a professional translator.",
    user_prompt_template="Translate this to Spanish: {text}",
    agent_type=AgentType.ONE_SHOT
)
class TranslatorAgent(Protocol):
    def translate(self, text: str) -> str: ...

2. Configure Credentials

Inject the centralized LLMConfig and set your keys. This avoids hardcoded environment variable lookups inside the library.

import os
from pico_ioc import component, configure
from pico_agent import LLMConfig

@component
class AppConfig:
    @configure
    def setup_llm(self, config: LLMConfig):
        # Load from Env, Vault, AWS Secrets, etc.
        config.api_keys["openai"] = os.getenv("OPENAI_API_KEY")
        config.api_keys["google"] = os.getenv("GOOGLE_API_KEY")

3. Run the Application

pico-agent includes its own init() that wraps pico_ioc.init() with automatic module inclusion and plugin discovery:

from pico_agent import init
from app.agents import TranslatorAgent

def main():
    # Scans modules, finds agents, wires dependencies
    container = init(modules=["app"])

    # Get your agent instance (auto-generated proxy)
    translator = container.get(TranslatorAgent)

    result = translator.translate(text="Dependency Injection is cool")
    print(result)

if __name__ == "__main__":
    main()

You can also use pico_ioc.init() directly — just include "pico_agent" in your modules list:

from pico_ioc import init

container = init(modules=["pico_agent", "app"])

Advanced Usage

ReAct Agents & Tools

Agents can use any dependency in the container as a tool.

# 1. Define a tool (standard component)
from pico_ioc import component

@component
class Calculator:
    def add(self, a: int, b: int) -> int:
        """Adds two numbers."""
        return a + b

# 2. Define the Agent
@agent(
    name="math_expert",
    capability=AgentCapability.REASONING,
    agent_type=AgentType.REACT,  # Enables Reasoning Loop
    tools=["calculator"],        # Reference tool by name (snake_case of class)
    max_iterations=5
)
class MathAgent(Protocol):
    def solve(self, problem: str) -> str: ...

Multi-Agent Orchestration

An agent can use other agents as tools simply by referencing them.

@agent(
    name="orchestrator",
    capability=AgentCapability.SMART,
    system_prompt="Coordinate the sub-agents to solve the task.",
    # Inject other agents as tools available to this LLM
    agents=["translator", "math_expert"] 
)
class Orchestrator(Protocol):
    def handle_request(self, task: str) -> str: ...

Model Routing & Configuration

pico-agent uses semantic capabilities to route to specific models. You can configure this globally using a configurer.

from pico_ioc import component, configure
from pico_agent.router import ModelRouter, AgentCapability

@component
class RouterConfig:
    @configure
    def configure_router(self, router: ModelRouter):
        # Update default mappings
        router.update_mapping(AgentCapability.FAST, "openai:gpt-5-mini")
        router.update_mapping(AgentCapability.SMART, "anthropic:claude-4-5-sonnet")
        
        # You can also use "provider:model" syntax
        router.update_mapping(AgentCapability.CODING, "deepseek:deepseek-coder")

With pico-boot

If you use pico-boot, pico-agent is automatically discovered via entry points — no need to include it in your modules list:

from pico_boot import init

container = init(modules=["app"])
# pico-agent is auto-discovered and loaded!

Testing

Testing is simple because you can mock the underlying LLMFactory or the agent protocol itself.

from unittest.mock import MagicMock
from pico_ioc import init
from pico_agent import LLMFactory, LLM

def test_my_agent():
    mock_llm = MagicMock(spec=LLM)
    mock_llm.invoke.return_value = "Mocked Response"
    
    mock_factory = MagicMock(spec=LLMFactory)
    mock_factory.create.return_value = mock_llm
    mock_factory.return_value = mock_factory # For singleton resolution

    container = init(
        modules=["pico_agent", "my_app"],
        overrides={LLMFactory: mock_factory},
    )
    
    agent = container.get(MyAgent)
    assert agent.run("test") == "Mocked Response"

AI Coding Skills

Install Claude Code or OpenAI Codex skills for AI-assisted development with pico-agent:

curl -sL https://raw.githubusercontent.com/dperezcabrera/pico-skills/main/install.sh | bash -s -- agent

All skills: curl -sL https://raw.githubusercontent.com/dperezcabrera/pico-skills/main/install.sh | bash

See pico-skills for details.

License

This project is licensed under the MIT License.