Agno Deep Agent

Agno Deep Agent is an opinionated harness for building deep agents with the Agno feel: easy to call, modular inside, and ready for long-running work.

It combines Team, Agent, Skills, SqliteDb, memory, learning, and workspace tools behind a small interface. The CLI also adds Agno-native context compression and multimodal attachments for models that support images, audio, video, or files:

from agno_deep_agents import create_deep_agent


agent = create_deep_agent(
    model="ollama:gemma4:e4b",
    instructions="Answer in a practical and concise way.",
)

agent.print_response("Analyze this project and propose the next step.", stream=True)

Goal

The project follows the Deep Agents idea: planning, subagents, filesystem, shell, memory, and persistent context. The difference is that the implementation uses Agno primitives instead of introducing a separate runtime.

Navigation

• Get Started
• CLI
• ACP Editors
• Python API
• Models and Providers
• Architecture
• Protocols
• Skills
• Roadmap
• Publish on GitHub Pages
• Troubleshooting

What Is Included

• TeamMode.tasks to coordinate discovery, implementation, and review.
• Specialist subagents: Researcher, Implementer, and Reviewer.
• SQLite at .deep-agent/agno.db for sessions, memory, and learning.
• Local skills in skills/.
• Workspace-restricted CodingTools.
• Agno context compression for long tool-heavy sessions.
• Multimodal inputs through Agno Image, Audio, Video, and File classes.
• Interactive Agno-colored CLI with /model, /models, /compress, and /attach.
• Support for Agno provider:model strings, including OpenAI, Anthropic, Google, Groq, local Ollama, and Ollama Cloud.
• Agno-inspired documentation styling with dark mode and syntax highlighting.
• ACP stdio server plus protocol direction for MCP and AgentOS.

Custom Tool Example

from agno_deep_agents import create_deep_agent


def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"


agent = create_deep_agent(
    model="ollama:gemma4:e4b",
    tools=[get_weather],
    instructions="You are a concise assistant.",
)

agent.print_response("What is the weather in sf?", stream=True)

Sandbox Note

Shell execution is restricted by workspace, command allowlist, timeout, and CodingTools operator checks. This reduces operational risk, but it is not the same as an operating-system or container sandbox.