Agent Definitions¶

Agent ¶

Bases: AgentMeta

The Agent class is a concrete implementation of an AI agent with tool-calling capabilities, inheriting from AgentMeta. It integrates a language model, tools, memory, and flow management to process queries, execute tools, and maintain conversational context.

Methods:

Name	Description
`ainvoke`
`invoke`
`stream`	Stream the agent response token-by-token with continuous tool-calling
`save_memory`	Save the tool message to the memory
`register_tools`	Register a list of tools.

ainvoke `async` ¶

ainvoke(query: str, is_save_memory: bool = False, user_id: str = 'unknown_user', max_iterations: int = 10, is_tool_formatted: bool = True, max_history: int = None, **kwargs) -> Awaitable[Any]

invoke ¶

invoke(query: str, is_save_memory: bool = False, user_id: str = 'unknown_user', max_iterations: int = 10, is_tool_formatted: bool = True, max_history: int = None, **kwargs) -> Any

stream ¶

stream(query: str, is_save_memory: bool = False, user_id: str = 'unknown_user', max_iterations: int = 10, is_tool_formatted: bool = True, max_history: int = None, **kwargs) -> Generator[Any, None, None]

Stream the agent response token-by-token with continuous tool-calling capability. Follows the same 3-step loop as :meth:invoke but yields AIMessageChunk objects so the caller can push tokens to a live connection (WebSocket, SSE, etc.) as soon as they arrive.

Workflow¶

Step 1 – Call the LLM with structured output to determine whether a tool is needed (AgentResponse).
If no tool needed: stream the direct answer token-by-token and return.
Step 2 – Execute the tool synchronously (same as invoke).
Step 3 – After all tool iterations, stream the final LLM summary token-by-token.

Parameters:

Name	Type	Description	Default
`query`	`str`	User query.	required
`is_save_memory`	`bool`	Save conversation to long-term memory.	`False`
`user_id`	`str`	Identifier for the current user.	`'unknown_user'`
`max_iterations`	`int`	Maximum tool-calling iterations.	`10`
`is_tool_formatted`	`bool`	If True, streams a final LLM summary after tool execution; if False, yields the raw ToolMessage.	`True`
`max_history`	`int`	Number of history messages to include.	`None`
`**kwargs`		Forwarded to the compiled-graph path when applicable.	`{}`

Yields:

Type	Description
`Any`	AIMessageChunk \| AIMessage \| ToolMessage: Streamed LLM tokens
`Any`	or the final tool/LLM message.

save_memory ¶

save_memory(message: Union[ToolMessage, AIMessage], user_id: str = 'unknown_user') -> None

Save the tool message to the memory

register_tools ¶

register_tools(tools: List[str]) -> Any

Register a list of tools.

Automatically detects whether each entry is an AgentSkill directory (contains a SKILL.md file) and routes it to :meth:~vinagent.register.tool.ToolManager.register_agentskill_tool. All other paths are treated as regular Python modules and forwarded to :meth:~vinagent.register.tool.ToolManager.register_module_tool.