Skip to main content
This is a major release that introduces a completely new approach to building multi-agent systems. It also introduces the AgentOS, a runtime for agents. This is a major rewrite of the Agno library and introduces various new concepts and updates to the existing ones. Some of the major changes are:
  • Agents, Teams and Workflows are now fully stateless.
  • Knowledge is now a single solution that supports many forms of content.
  • Storage of sessions, memories, evals, etc. has been simplified

​
General Changes

  • /libs/agno has been restructured to fit the new concepts in Agno and for better organization.
  • All code related to managing workspaces and agent deployment in Agno has been moved to a new package called agno-infra. This is a combination of the previous agno-aws and agno-docker packages, as well as the CLI and other tools.
  • agno-aws and agno-docker packages have been deprecated and will no-longer be maintained.
  • All code related to the Agno CLI (ag) has been moved to this new agno-infra package.
  • Added AgentOS to agno as a comprehensive API solution for building multi-agent systems. This also replaces Playground and other Apps. See details below.
  • Cookbook has been completely restructured, with new and more valuable READMEs, better coverage of concepts, and more examples.
  • Introducing AgentOS, a system for hosting agents, teams and workflows as a production-ready API. See full details in the AgentOS section.
  • This adds routes for session management, memory management, knowledge management, evals management, and metrics.
  • This enables you to host agents, teams and workflows, and use the Agent OS UI to manage them.
  • Removed Playground. Its functionality has been replaced by AgentOS.
  • Removed AGUIApp and replace with AGUI interface on AgentOS.
  • Removed SlackApi and replace with Slack interface on AgentOS.
  • Removed WhatsappApi and replace with Whatsapp interface on AgentOS.
  • Removed FastAPIApp. Its functionality has been replaced by AgentOS.
  • DiscordClient has been moved to /integrations/discord.

​
Session & Run State

  • We have made significant changes to the innerworkings of Agent, Team and Workflow to make them completely stateless.
  • This means that agent_session, session_metrics, session_state, etc. should not be seen as stateful variables that would be updated during the course of a run, but rather as “defaults” for the agent if they can be set on initialisation.
  • CustomEvent is now supported and you can inherit from it to create your own custom events that can be yielded from your own tools. See the documentation for more details.
For agents:
  • RunResponse -> RunOutput
  • RunResponseStartedEvent -> RunStartedEvent
  • RunResponseContentEvent -> RunContentEvent
  • RunResponseCompletedEvent -> RunCompletedEvent
  • IntermediateRunResponseContentEvent -> IntermediateRunContentEvent
  • RunResponseErrorEvent -> RunErrorEvent
  • RunResponseCancelledEvent -> RunCancelledEvent
For teams:
  • TeamRunResponse -> TeamRunOutput
  • RunResponseStartedEvent -> RunStartedEvent
  • RunResponseContentEvent -> RunContentEvent
  • RunResponseCompletedEvent -> RunCompletedEvent
  • IntermediateRunResponseContentEvent -> IntermediateRunContentEvent
  • RunResponseErrorEvent -> RunErrorEvent
  • RunResponseCancelledEvent -> RunCancelledEvent
For workflows:
  • WorkflowRunResponse -> WorkflowRunOutput
  • WorkflowRunResponseStartedEvent -> WorkflowRunStartedEvent
  • WorkflowRunResponseContentEvent -> WorkflowRunContentEvent
  • WorkflowRunResponseCompletedEvent -> WorkflowRunCompletedEvent
  • WorkflowRunResponseErrorEvent -> WorkflowRunErrorEvent
  • WorkflowRunResponseCancelledEvent -> WorkflowRunCancelledEvent
  • The import location for RunOutput (and events) has been moved to agno.run.agent.
  • For RunOutput, TeamRunOutput and WorkflowRunOutput the extra_data attribute has been removed and the internal attributes are now top-level. This is references, additional_input, reasoning_steps, and reasoning_messages.
  • metadata added to RunOutput, TeamRunOutput and WorkflowRunOutput. This represents all the set metadata for the run.
  • Session storage now stores AgentSession, TeamSession and WorkflowSession with new schemas. See full details in the Session section.
  • Session objects now have runs directly on it.
  • Session objects support new convenience methods:
    • get_run -> Get a specific run by ID.
    • get_session_summary -> Get the session summary.
    • get_chat_history -> Get an aggregated view of all messages for all runs in the session.
  • SessionMetrics and MessageMetrics have been unified as a single Metrics class.
  • audio_tokens has been renamed to audio_total_tokens.
  • input_audio_tokens has been renamed to audio_input_tokens.
  • output_audio_tokens has been renamed to audio_output_tokens.
  • cached_tokens has been renamed to cache_read_tokens.
  • prompt_tokens and completion_tokens have been removed (only input_tokens and output_tokens should be used)
  • prompt_tokens_details and completion_tokens_details have been removed. Instead provider_metrics captures any provider-specific metrics.
  • time has been renamed to duration.
  • You can now cancel a run by calling cancel_run on the Agent, Team or Workflow.
  • This will cancel the run and return a RunCancelledEvent during streaming, or set the RunOutput.status to "cancelled".

​
Storage

  • Agent, Team, Workflow and the various evals now all support a single db parameter. This is to enable storage for the instance of that class. This is required for persistence of sessions, memories, metrics, etc.
  • storage and memory have been removed from Agent, Team and Workflow.
  • This means all previous storage providers have been reworked. Also session storage, memory storage and eval storage are all a single solution now referred to as a “DB”.
  • PostgresStorage -> PostgresDb
  • SqliteStorage -> SqliteDb
  • MysqlStorage -> MysqlDb
  • RedisStorage -> RedisDb
  • MongoStorage -> MongoDb
  • DynamoDBStorage -> DynamoDb
  • SingleStoreStorage -> SingleStoreDb
  • InMemoryStorage -> InMemoryDb
  • JsonStorage -> JsonDb
  • GCSJsonStorage -> GCSJsonDb

​
Memory

  • With the above changes to storage, memory has been simplified.
  • memory has been removed from Agent and Team. Instead memory is enabled with enable_user_memories: bool (like before) and persisted in the db instance.
  • Changes to how memories are created can still be done by overriding the MemoryManager class on Agent or Team. E.g. Agent(memory_manager=MyMemoryManager()).
  • AgentMemory and TeamMemory have been removed.

​
Knowledge

  • Knowledge has been completely reworked. See full details in the Knowledge section.
  • You now define a single Knowledge instance for all types of content. Files (PDF, CSV, etc.), URLs, and other.
  • The agent can still use your knowledge base to search for information at runtime. All existing RAG implementations are still supported.
  • Added full async support for embedding models and vector DBs. This has a significant impact on performance and is a major speed improvement when adding content to the knowledge base using knowledge.add_content_async(...).
  • AgentKnowledge and all other knowledge base classes have been removed.
  • Import locations for embedder, document, chunking, reranker and reader have been moved to agno.knowledge. See examples for more details.

​
Tools updates

  • General:
    • Since Agents and Teams are now stateless, using attributes from the agent/team object inside a function will give you access to the attributes set on initialisation of that agent/team. E.g. agent.session_state should not be used, instead session_state can now be directly accessed and would have the “current” state of the session.
    • A new flow allows images, audio and video files generated during tool execution to be passed back in a FunctionExecutionResult object and this will ensure these artifacts are made available to the model and agent as needed.
  • All tools that handle media (e.g. image generation tools) now correctly add this generated media to the RunOutput, but also make it available for subsequent model calls.
  • The interface of almost all the toolkits have been updated for a more consistent experience around switching specific tools on and off. The list of changes is too long to list here. We suggest you take a look at the toolkits you use specifically and how they have been updated.
  • show_results is now True by default for all tools. If you just set stop_after_tool_call=True then show_results will be automatically set to True.
  • images, videos, audio and files are now available as parameters to tools. See the documentation for more details.

​
Media

​
Removals

  • Removed legacy artifact classes: ImageArtifact, VideoArtifact, AudioArtifact, and AudioResponse classes have been completely removed in favor of unified media classes.

​
New Unified Media Architecture

  • Unified Image class: Now serves all use cases (input, output, artifacts) with standardized content: Optional[bytes] field for raw image data
  • Unified Audio class: Replaces both AudioArtifact and AudioResponse with consistent byte-based content storage and additional fields like transcript, expires_at, sample_rate, and channels
  • Unified Video class: Updated to handle all video use cases with standardized content handling and metadata fields
  • Enhanced File class: Updated to work seamlessly across agent, team, workflow, and toolkit contexts

​
New Methods and Features

  • from_base64() class method: Added to Image, Audio, and Video classes for creating instances from base64-encoded content (automatically converts to raw bytes)
  • get_content_bytes() method: Retrieves content as raw bytes, handling loading from URLs or file paths
  • to_base64() method: Converts content to base64 string for transmission/storage
  • to_dict() method: Enhanced serialization with optional base64 content inclusion

​
Content Standardization

  • Byte-based storage: All media content is now stored as raw bytes (Optional[bytes]) instead of mixed string/bytes formats
  • Automatic validation: Model validators ensure exactly one content source (url, filepath, or content) is provided
  • Auto-generated IDs: Media objects automatically generate UUIDs when not provided

​
Logging

  • Added support for custom loggers. See the documentation for more details.

​
Agent updates

  • agent_id -> id -> If id is not set, it is autogenerated using the name of the agent, or a random UUID if the name is not set.
  • search_previous_sessions_history -> search_session_history
  • context -> dependencies
  • add_context -> add_dependencies_to_context
  • add_history_to_messages -> add_history_to_context
  • add_name_to_instructions -> add_name_to_context
  • add_datetime_to_instructions -> add_datetime_to_context
  • add_location_to_instructions -> add_location_to_context
  • add_messages -> additional_input
  • extra_data -> metadata
  • create_default_system_message -> build_context
  • create_default_user_message -> build_user_context
  • Added send_media_to_model -> True by default. Set to False if you don’t want to send media (image, audio, video, files) to the model. This is useful if you only want media for tools.
  • Added store_media -> True by default. Set to False if you don’t want to store any media in the RunOutput that is persisted with sessions.
  • num_history_responses -> num_history_runs
  • Removed success_criteria and goal
  • Removed team_session_id and workflow_session_id.
  • Removed introduction
  • Removed show_tool_calls -> This is now just always enabled.
  • Removed team and team_data
  • Removed respond_directly, add_transfer_instructions, team_response_separator and team_session_id (since team has been removed from Agent)
  • Removed all team functionality from inside Agent (i.e. the deprecated teams implementation has been removed)
  • Removed all monitoring from Agent. With the new AgentOS platform, monitoring is done using your own data. Go to os.agno.com to get started.
  • response_model -> output_schema
  • Added input_schema (a pydantic model) to validate the input to the agent.
  • Changed message to input (which also replaces messages). input can be of type str, list, dict, Message, BaseModel, or list[Message].
  • If a dict and input_schema is provided, the dict will be validated against the schema.
  • If a BaseModel and input_schema is provided, the model will be validated against the schema.
  • arun and acontinue_run with stream=True now return an async iterator of RunOutputEvent directly and is not a coroutine anymore.
  • debug_mode: bool added to run, arun, print_response and aprint_response to enable debug mode for a specific run.
  • add_history_to_context added to run, arun, print_response and aprint_response to add the chat history to the context for the current run.
  • dependencies added to run, arun, print_response and aprint_response to add dependencies to the context for the current run.
  • metadata added to run, arun, print_response and aprint_response to set the metadata for the current run. This is merged with the metadata set on the Team object.
  • Added get_run_output and get_last_run_output to Agent to retrieve a run output by ID.
  • Metrics have been simplified and cleaned up.
  • There are now 3 levels of metrics:
    • Message.metrics -> Metrics for each message (assistant, tool, etc.).
    • RunOutput.metrics -> Aggregated metrics for the whole run.
    • AgentSession.metrics -> Aggregated metrics for the whole session.
  • knowledge is now an instance of Knowledge instead of AgentKnowledge.
  • retriever -> knowledge_retriever -> For a custom retriever.
  • add_references -> add_knowledge_to_context -> To enable traditional RAG.
  • add_memory_references -> add_memories_to_context
  • You can set a custom memory_manager to use when creating memories.
  • Added get_user_memories to retrieve the user memories.
  • add_session_summary_references -> add_session_summary_to_context
  • You can set a custom session_summary_manager to use when creating session summaries.
  • Removed session_name and replace with functions get_session_name and rename_session.
  • Added get_session to retrieve a session by ID.
  • Added get_chat_history to retrieve the chat history from a session.
  • Added get_session_metrics to retrieve the metrics for a session.
  • Added get_session_state to retrieve the session state from a session.
  • Added get_session_summary to retrieve the session summary from a session.
  • Because Agent is now stateless, agent_session, session_metrics, run_id, run_input, run_messages and run_response as “sticky” agent attributes have been removed.
  • Because Agent is now stateless, images, videos, audio are no longer available as agent attributes. Instead these can be accessed on the RunOutput for a particular run.
  • Removed team_session_state and workflow_session_state. Only session_state is used.
  • Added enable_agentic_state to Agent and Team to allow the agent to update the session state with a tool call.

​
Team updates

  • Removed mode from Team. Instead there are attributes that can be used to control the behavior of the team:
    • respond_directly -> If True, the team leader won’t process responses from the members and instead will return them directly
    • delegate_task_to_all_members -> If True, the team leader will delegate tasks to all members simultaneously, instead of one by one. When running async (using arun) members will run concurrently.
    • determine_input_for_members -> True by default. Set to False if you want to send the run input directly to the member agents without the team leader synthesizing its own input.
  • team_id -> id -> If id is not set, it is autogenerated using the name of the team, or a random UUID if the name is not set.
  • search_previous_sessions_history -> search_session_history
  • context -> dependencies
  • add_context -> add_dependencies_to_context
  • add_history_to_messages -> add_history_to_context
  • add_name_to_instructions -> add_name_to_context
  • add_datetime_to_instructions -> add_datetime_to_context
  • add_location_to_instructions -> add_location_to_context
  • add_member_tools_to_system_message -> add_member_tools_to_context
  • extra_data -> metadata
  • Added additional_input (works the same as for Agent)
  • Added store_member_responses: bool to optionally store the member responses on the team run output object.
  • Added acli_app to Team to enable the CLI app for the team in async mode.
  • Added send_media_to_model -> True by default. Set to False if you don’t want to send media (image, audio, video, files) to the model. This is useful if you only want media for tools.
  • Added store_media -> True by default. Set to False if you don’t want to store any media in the RunOutput that is persisted with sessions.
  • num_history_responses -> num_history_runs
  • Removed success_criteria
  • Removed team_session_id and workflow_session_id.
  • Removed enable_team_history
  • Removed num_of_interactions_from_history
  • Removed show_tool_calls -> This is now just always enabled.
  • Removed enable_agentic_context. session_state and enable_agentic_state should rather be used to manage state shared between the team and the members.
  • Removed all monitoring from Team. With the new AgentOS platform, monitoring is done using your own data. Go to os.agno.com to get started.
  • response_model -> output_schema
  • Added input_schema (a pydantic model) to validate the input to the agent.
  • Changed message to input (which also replaces messages). input can be of type str, list, dict, Message, BaseModel, or list[Message].
  • If a dict and input_schema is provided, the dict will be validated against the schema.
  • If a BaseModel and input_schema is provided, the model will be validated against the schema.
  • arun with stream=True now return an async iterator of TeamRunOutputEvent directly and is not a coroutine anymore.
  • debug_mode: bool added to run, arun, print_response and aprint_response to enable debug mode for a specific run.
  • add_history_to_context added to run, arun, print_response and aprint_response to add the chat history to the context for the current run.
  • dependencies added to run, arun, print_response and aprint_response to add dependencies to the context for the current run.
  • metadata added to run, arun, print_response and aprint_response to set the metadata for the current run. This is merged with the metadata set on the Team object.
  • Added get_run_output and get_last_run_output to Team to retrieve a run output by ID.
  • Metrics have been simplified and cleaned up.
  • There are now 3 levels of metrics:
    • Message.metrics -> Metrics for each message (assistant, tool, etc.).
    • RunOutput.metrics -> Aggregated metrics for the whole run.
    • TeamSession.metrics -> Aggregated metrics for the whole session.
  • knowledge is now an instance of Knowledge instead of AgentKnowledge.
  • retriever -> knowledge_retriever -> For a custom retriever.
  • add_references -> add_knowledge_to_context -> To enable traditional RAG.
  • Added update_knowledge tool to update the knowledge base. Works the same as for Agent.
  • add_memory_references -> add_memories_to_context
  • You can set a custom memory_manager to use when creating memories.
  • Added get_user_memories to retrieve the user memories.
  • add_session_summary_references -> add_session_summary_to_context
  • You can set a custom session_summary_manager to use when creating session summaries.
  • Removed session_name and replace with functions get_session_name and rename_session.
  • Added get_session to retrieve a session by ID.
  • Added get_chat_history to retrieve the chat history from a session.
  • Added get_session_metrics to retrieve the metrics for a session.
  • Added get_session_state to retrieve the session state from a session.
  • Added get_session_summary to retrieve the session summary from a session.
  • Because Team is now stateless, team_session, session_metrics, run_id, run_input, run_messages and run_response as “sticky” team attributes have been removed.
  • Because Team is now stateless, images, videos, audio are no longer available as team attributes. Instead these can be accessed on the TeamRunOutput for a particular run.
  • Removed team_session_state and workflow_session_state. Only session_state is used.
  • Added enable_agentic_state to Team to allow the agent to update the session state with a tool call.

​
Workflow updates

  • workflow_id -> id -> If id is not set, it is autogenerated using the name of the workflow, or a random UUID if the name is not set.
  • Workflows “v1” has been completely removed and replaced with Workflows v2. See full details in the Workflows section.
  • This means the import locations for “Workflows v2” is now agno.workflows.
  • extra_data -> metadata
  • Added store_events to Workflow to optionally store the events on the workflow run output object. Also added events_to_skip to skip certain events from being stored. This works the same as for Agent and Team.
  • Added store_executor_outputs to Workflow to optionally store the agent/team responses on the workflow run output object.
  • Added input_schema to Workflow to validate the input to the workflow.
  • Added support for websocket streaming of the workflow. This is appropriate for long-running workflows that need to be streamed to a client. This is only available for arun.
  • Removed all monitoring from Workflow. With the new AgentOS platform, monitoring is done using your own data. Go to os.agno.com to get started.
  • Changed message to input (which also replaces messages). input can be of type str, list, dict, or BaseModel.
  • If a dict and input_schema is provided, the dict will be validated against the schema.
  • If a BaseModel and input_schema is provided, the model will be validated against the schema.
  • arun with stream=True now return an async iterator of WorkflowRunOutputEvent directly and is not a coroutine anymore.
  • debug_mode: bool added to run, arun, print_response and aprint_response to enable debug mode for a specific run.
  • Added get_run_output and get_last_run_output to Workflow to retrieve a run output by ID.
  • Removed session_name and replace with functions get_session_name and rename_session.
  • Because Workflow is now stateless, workflow_session, session_metrics, run_id, run_input, run_messages and run_response as “sticky” workflow attributes have been removed.
  • Because Workflow is now stateless, images, videos, audio are no longer available as workflow attributes. Instead these can be accessed on the WorkflowRunOutput for a particular run.
  • Added get_session to retrieve a session by ID.
  • Added get_session_metrics to retrieve the metrics for a session.
  • Added get_session_state to retrieve the session state from a session.
  • Added get_session_summary to retrieve the session summary from a session.
⌘I