Agno
also known as PhiData
Type: full-code · Vendor: Agno · Language: Python · License: Apache-2.0 · Status: active · Status in practice: mature · First released: 2022-05-04
Provide an SDK and runtime for building agent platforms — role-defined Agents, multi-mode Teams (coordinate / route / broadcast / sequential), Workflows, persistent memory/storage/knowledge, MCP toolkits, and AgentOS with cron scheduling, RBAC, OpenTelemetry tracing, and human-approval gates.
Description. Agno (formerly PhiData) is an Apache-2.0 Python SDK that frames the world as Agents — 'AI programs that use tools to accomplish tasks'. On top of agents it adds Teams with coordination modes (Coordinate / Route / Broadcast / Sequential), step-based Workflows, persistent Memory with a MemoryManager for LLM-driven user-memory extraction, Agentic-RAG via the Knowledge module, and 100+ integration Toolkits including MCPTools (stdio / SSE / streamable HTTP). The companion AgentOS platform exposes the runtime and a control plane for scheduled jobs, RBAC, OTel tracing, and human-approval gates.
Agent loop shape. Stateful tool-calling loop. Each Agent run assembles instructions, knowledge (RAG), memory, and tool schemas; the model emits tool calls or a final reply; tool results are appended and the loop iterates until the model returns text or hits a cap. Team wraps this loop with a leader+members topology selected via mode: Coordinate (leader picks members, crafts tasks, synthesises), Route (leader routes to one specialist), Broadcast (same task to all members), Sequential (chain agent responses). With persistent memory enabled the agent stores user memories across runs, sessions, and agents.
Primary use cases
- multi-agent teams with Coordinate / Route / Broadcast / Sequential modes
- Agentic RAG via the Knowledge module
- persistent user memory across runs and sessions
- production agent platform via AgentOS (cron + RBAC + OTel)
- MCP-integrated agents with 100+ toolkit integrations
Key concepts
- Agent (docs) — AI program that uses tools to accomplish tasks; the primary unit.
- Team modes → orchestrator-workers (docs) — Coordinate / Route / Broadcast / Sequential — four documented multi-agent orchestration shapes.
- Workflow (docs) — Step-based sequential and parallel execution wrapping agents, with conditional branches.
- Memory + MemoryManager → cross-session-memory (docs) — User preferences and history persisted across conversations; MemoryManager drives LLM-based memory extraction.
- Knowledge / Agentic RAG → agentic-rag (docs) — Searchable knowledge bases that the agent can autonomously retrieve from.
- MCPTools → mcp (docs) — Toolkit that integrates MCP servers via stdio / SSE / streamable HTTP.
- AgentOS (docs) — The runtime and control plane that runs your agent platform.
Patterns this full-code implements —
- ★★Agentic RAG
Knowledge module powers autonomous retrieval; documented as 'Agentic RAG'.
- ★★Cross-Session Memory
MemoryManager persists user memories across runs/sessions/agents.
- ★★Model Context Protocol
MCPTools and MultiMCPTools speak stdio / SSE / streamable HTTP; AgentOS can also be exposed as an MCP server.
- ★★Orchestrator-Workers
Team modes include Coordinate (leader picks members, crafts tasks, synthesises) and Broadcast (same task to all members, synthesise).
- ★★Supervisor
Router pattern: leader directs tasks to a specific specialist; Sequential pattern chains member responses.
- ★★ReAct
Agent run loop iterates model + tool calls until a final message without tool calls is produced.
- ★★Tool Use
Tools are functions agents call to interact with external systems; 120+ pre-built toolkits ship with the framework.
- ★★Structured Output
Agents accept Pydantic models via output_schema and return validated typed objects rather than text.
- ★★Human-in-the-Loop
@tool(requires_confirmation=True) marks tools that pause the run for user approval; AgentOS surfaces pending tool calls for admin approval.
- ★★Scheduled Agent
AgentOS Scheduler runs agents, teams, and workflows on cron schedules with configurable intervals, retries, and timezones.
- ★★Code Execution
PythonTools and ShellTools are pre-built toolkits that let agents execute Python and shell commands; save_to_file_and_run / run_python_code are explicit code-execution functions.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Instantiates
Alternatives & relatives
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified