Agno

Agno has three building blocks. An Agent builds context for a model, runs the tools it requests and returns a run output; memory, knowledge and storage are optional add-ons. A Team has a leader model and members (agents or nested teams); its mode decides whether the leader decomposes and delegates tasks (coordinate, the default), hands the request to one member (route), sends it to all members (broadcast) or works through a shared task list (tasks). A Workflow runs agents, teams and plain functions as steps with sequential, parallel, loop, condition and router control flow.

AgentOS is the runtime. It serves agents, teams and workflows as a FastAPI application with streaming endpoints, and persists sessions, runs, memory, traces, schedules and approvals through a shared db interface (Postgres, SQLite and other backends). Interfaces attach the same components to chat apps and to A2A, AG-UI and MCP clients, and remote-execution helpers call agents on other AgentOS instances.

A browser-based control plane connects to local or deployed AgentOS runtimes to inspect traces and sessions, edit memory and knowledge, resolve approvals and manage schedules; the data stays in the runtime's database.

Marking a tool with @tool(requires_confirmation=True) pauses the run before the call: the run output has is_paused set, the application approves or rejects the pending tool call, and continue_run() resumes. Other documented modes collect specific user input, let the agent ask for input when it needs it, hand a tool call to external code for execution, or persist an approval for admin review. With a database configured, a paused run can be continued later by run_id and session_id, even from a new process. Workflows can pause before or after a step or at a tool call inside a step; pause state is saved to the workflow database and workflow.continue_run() picks up from there. Approvals can also be resolved in the AgentOS control plane.

pip install -U agno openai

from agno.agent import Agent
from agno.models.openai import OpenAIResponses
from agno.team import Team
from agno.team.mode import TeamMode

model = OpenAIResponses(id="gpt-5.4-mini")  # reads OPENAI_API_KEY
billing = Agent(name="Billing Agent", role="Answer billing and refund questions", model=model)
tech = Agent(name="Tech Agent", role="Answer setup and troubleshooting questions", model=model)

helpdesk = Team(
    name="Helpdesk",
    model=model,
    members=[billing, tech],
    mode=TeamMode.route,  # the leader picks one member, who answers directly
)
helpdesk.print_response("My invoice shows two charges for March.")

• Package metadata requires Python >=3.9; the setup guide creates a Python 3.12 environment.
• Model providers are separate installs (openai here) and need their API key, such as OPENAI_API_KEY.
• Features live behind extras: agno[os] for AgentOS, agno[mcp], agno[os,a2a], agno[os,agui].
• Agno sends a telemetry event per agent run; set AGNO_TELEMETRY=false to disable it.
• Upgrading from v2 to v3 requires a database migration (runs move to their own table) and removing the retired Culture feature.
• To resume a paused HITL run in another process, keep both run_id and session_id and use the same database.

Yes, in both directions. MCPTools connects agents to MCP servers, and AgentOS can publish agents, teams and workflows as an MCP server for clients such as Claude Code, Codex and Cursor.

Yes. AgentOS ships A2A and AGUI interfaces, and A2AClient calls other A2A servers. Both interfaces need their install extras.

The SDK is Apache-2.0 and the runtime is free to self-host. The pricing page lists a free control plane for local runtimes, a $150/month Pro plan for live runtimes, and custom Enterprise pricing.

Yes. ClaudeAgent wraps the Claude Agent SDK so Claude Code can be served through AgentOS with tool permission rules.

It sets the orchestration: coordinate (leader delegates and synthesizes), route (one member answers), broadcast (all members get the task) or tasks (a shared task list).