AllegroAgent

A lightweight, extensible Python framework for building stateful AI agents backed by local or remote LLMs. Agent is the single public entry point — configure it with a model string and a list of tools, and call run(). The framework handles conversation history, provider routing, and the tool-calling loop for you.

---

Features

• Stateful agents with tool use — Agent.run() manages conversation history and automatically executes tools the LLM requests, looping until it produces a final answer.
• Resumable conversations — pass prior messages via Agent(history=...) to pick up where another session left off; export the current state any time with agent.history().
• Pluggable providers — add new backends by implementing a single generate() method. Ollama ships in the box.
• JSON Schema tools — subclass BaseTool, declare parameters, and the agent exposes them to any function-calling-capable model.
• Typed errors — a small hierarchy of framework exceptions for clean error handling.
• Zero heavy deps — just requests at runtime.

---

Architecture

┌─────────────────────────────────────────────┐
│  Agent            (public, stateful loop)   │
├─────────────────────────────────────────────┤
│  Providers        (Ollama, …)               │
└─────────────────────────────────────────────┘

Agent talks to providers through a small internal router (_Client) that parses the provider:model string and caches provider instances. You never need to touch it directly.

allegro_agent/
├── agent.py              # Agent — the public entry point
├── _client.py            # internal — provider router used by Agent
├── exceptions.py         # FrameworkError hierarchy
├── providers/
│   ├── base.py           # BaseProvider + ProviderResponse dataclass
│   ├── ollama.py         # OllamaProvider
│   └── __init__.py       # Provider registry (get/register_provider)
└── tools/
    ├── base.py           # BaseTool with JSON Schema + to_schema()
    └── file_write.py     # FileWriteTool (reference implementation)

---

Installation

Requires Python 3.10+.

Install from PyPI:

pip install allegro-agent

Or install from source (for development):

git clone https://github.com/ajithraghavan/AllegroAgent.git
cd AllegroAgent
pip install -e ".[dev]"

For the default Ollama provider, install and start Ollama locally:

ollama serve
ollama pull llama3

---

Quick Start

from allegro_agent import Agent, FileWriteTool

agent = Agent(
    name="Writer",
    model="ollama:llama3",
    temperature=0.1,
    system_prompt="You are a helpful writing assistant.",
    tools=[FileWriteTool()],
)

print(agent.run("Write a haiku about the ocean to ocean.txt"))
print(agent.run("Now write one about mountains to mountains.txt"))
agent.reset()  # clear history

What happens on each run():

1. The agent sends the prompt + your tool schemas to the LLM.
2. If the LLM returns tool_calls, the agent executes each tool and feeds the results back.
3. Steps 1–2 repeat until the LLM returns a plain text answer (or MAX_TOOL_ROUNDS = 10).

---

Resuming a conversation

Agent keeps conversation state in memory only. To carry a chat across processes, requests, or storage backends of your choice, capture it with agent.history() and seed a new Agent with history=....

import json
from allegro_agent import Agent

# Session 1 — have a conversation, then capture state.
agent = Agent(model="ollama:llama3", system_prompt="You are helpful.")
agent.run("My name is John.")
agent.run("Remember that.")

snapshot = agent.history()                   # list[dict], JSON-serializable
json.dump(snapshot, open("conv.json", "w"))  # caller picks the storage

# Session 2 — possibly in a fresh process — resume.
prior = json.load(open("conv.json"))
agent = Agent(model="ollama:llama3", system_prompt="You are helpful.", history=prior)
print(agent.run("What is my name?"))         # → references "John"

Notes:

• history() returns a deep copy, so mutating the returned list never affects the agent.
• history is validated on construction. Each item must be a dict with role (one of "user", "assistant", "tool") and content (str). Malformed payloads raise InvalidHistoryError.
• The "system" role is not stored in history — system prompts continue to flow through the existing system_prompt kwarg, which the caller re-supplies (or omits) when constructing the resumed agent.
• The framework owns no I/O. Use JSON, a database, Redis, or anything else — it's just a list of plain dicts.

---

Configuration

Agent accepts the following keyword arguments (all keyword-only):

Argument	Type	Description
model	str	Required. Format: provider:model
name	str	Display name (default "Agent")
temperature	float	Sampling temperature
max_tokens	int	Max response tokens
system_prompt	str	System instruction prepended to every call
tools	list[BaseTool]	Tools the agent may invoke
history	list[dict]	Prior conversation messages to resume from

---

Extending the Framework

Adding a new provider

Subclass BaseProvider and register it:

from allegro_agent.providers.base import BaseProvider, ProviderResponse
from allegro_agent import register_provider

class OpenAIProvider(BaseProvider):
    def generate(self, messages, **kwargs) -> ProviderResponse:
        # call the API, translate the response
        return ProviderResponse(
            content="...",
            model=kwargs["model"],
            provider="openai",
            tool_calls=None,
        )

register_provider("openai", OpenAIProvider)
# now usable as model="openai:gpt-4"

Adding a new tool

Subclass BaseTool, define JSON Schema parameters, implement execute:

from allegro_agent import BaseTool

class AddTool(BaseTool):
    name = "add"
    description = "Add two integers."
    parameters = {
        "type": "object",
        "properties": {
            "a": {"type": "integer"},
            "b": {"type": "integer"},
        },
        "required": ["a", "b"],
    }

    def execute(self, **kwargs) -> str:
        return str(kwargs["a"] + kwargs["b"])

Pass an instance via the agent's tools list — the framework calls to_schema() to expose it to the LLM.

---

Exceptions

All framework errors inherit from FrameworkError:

• ProviderError — provider-level failures (network, API errors)
• ProviderNotFoundError — unknown provider name in the registry
• InvalidModelFormatError — model string missing provider: prefix
• ToolError — raised by tools on execution failure
• InvalidHistoryError — history argument to Agent is malformed

---

Testing

pytest

The reference end-to-end script in tests/test_agent.py exercises Agent against a live Ollama instance.

---

Project Metadata

• Package: allegro-agent v0.3.0
• Python: ≥ 3.10
• Runtime deps: requests
• License: see LICENSE