Library of the Week — Agno

Agno — lightweight, model-agnostic framework for building typed, multi-modal AI agents

GitHub · Language: Python · License: Mozilla Public License 2.0

What it does

Agno (formerly Phidata) is a framework for composing production-grade AI agents with built-in support for tools, memory, knowledge bases, and structured outputs. It targets developers who want the agent orchestration primitives of LangGraph or LlamaIndex without the abstraction overhead — you get plain Python classes that map directly to concepts you already understand.

Why it stands out

• Model-agnostic with first-class support for current flagships — Agno ships providers for OpenAI (GPT-5.6 Sol/Terra/Luna tiers), Anthropic (Claude Fable 5, Opus 5), Google (Gemini 3.7 Flash), and local models via Ollama, all behind a unified interface with no adapter boilerplate
• Typed tool definitions by default — tools are regular Python functions decorated with @tool; Agno extracts the schema from type hints and docstrings automatically, so you get structured calls without reaching for Instructor or writing JSON Schema by hand
• Built-in memory and storage backends — persistent agent memory via SQLite, PostgreSQL, or any custom store is a first-class feature, not a plugin — useful when building stateful assistants without pulling in a separate memory layer
• Multi-agent teams with a simple API — you compose a Team of Agent objects and route tasks with a coordinator; the communication model is transparent Python objects rather than compiled graphs, making debugging substantially easier than LangGraph’s state machine approach

Quick start

from agno.agent import Agent
from agno.models.openai import OpenAIChat
from agno.tools.duckduckgo import DuckDuckGoTools

agent = Agent(
    model=OpenAIChat(id="gpt-5.6-sol"),
    tools=[DuckDuckGoTools()],
    description="Research assistant that cites sources.",
    markdown=True,
)

agent.print_response("What shipped at Google I/O 2026?", stream=True)

When to use it

• You want a production-oriented agent framework that stays close to Python idioms — no custom DSLs, no YAML pipeline configs, no compiled computation graphs
• You’re building stateful assistants or multi-step research agents that need persistent memory and tool use without assembling five separate libraries
• You want to swap models between OpenAI, Anthropic, and Gemini tiers mid-project without rewriting agent logic

When to skip it

• If your primary need is complex branching workflows with fine-grained conditional routing, LangGraph’s explicit state machine is still better suited — Agno’s team coordinator is simpler but less expressive for deeply conditional pipelines
• If you’re doing pure retrieval-augmented generation at scale, LlamaIndex’s retrieval pipeline primitives are more mature

The verdict

Agno occupies the productive middle ground between bare SDK calls and heavyweight orchestration frameworks — you get real agent infrastructure (memory, tools, multi-agent routing) with a codebase you can actually read. It has seen a significant uptick in production adoption as teams look for alternatives to LangChain’s abstraction depth, and the model-provider breadth is genuinely competitive today. If you’re starting a new agent project and don’t already have strong opinions about your orchestration layer, Agno is a reasonable default.