Skip to main content
The Agent class wraps the VectorShift agent platform: declarative tools, typed inputs/outputs, and multi-turn conversations via Session.
Prerequisites: Installed SDK · API key set · Python 3.10+.

​
Mental model

  • An agent is an LLM + tools + instructions + typed I/O.
  • Two flavours: Functional (single call, you get outputs back — designed to be invoked from inside a Pipeline) and Conversational (multi-turn, runs behind a Session). Pick with type=AgentType.FUNCTIONAL or AgentType.CONVERSATIONAL.
  • Tools are declared with Tool + ToolInput. ~200 integration tools are auto-generated under vectorshift.agent.tools.
  • Every method has an async variant (anew, arun, acreate_session, …).
Where functional agents live. A functional agent is most useful as a node inside a Pipeline — drop it into a graph the same way you’d drop an LLM node, and it handles tool selection autonomously while the surrounding pipeline handles wiring, branching, and orchestration. Standalone agent.run({...}) is fine for one-off calls; pipeline-embedded is the common production shape.

​
Quick start

from vectorshift import Agent, LlmInfo

agent = Agent.new(
    name="Researcher",
    llm_info=LlmInfo(provider="openai", model_id="gpt-5.1"),
    instructions="You are a helpful research assistant.",
)
agent.save()

result = agent.run({"input": "What is the population of Tokyo?"})
print(result.output)
Sessions require an async context. If you paste a Session snippet into a .py file, wrap it in async def main(): … and call asyncio.run(main()).

​
Functional vs Conversational

FunctionalConversational
ShapeOne call, one resultMulti-turn, streaming events
Entry pointagent.run({...}) — typically called from inside a Pipeline nodeagent.create_session() then session.send() / session.listen()
Use whenYou want autonomous tool selection as one step of a larger pipeline (extraction, classification, summarisation, multi-tool research)You’re building a chatbot, want token-level streaming, or need human-in-the-loop approvals
StateStateless — each call is independentStateful per session_id (memory carries across turns when MemoryConfig.enable_session_memory=True)

​
Recent additions

The recent agent revamp introduced the Tool / ToolInput system, an auto-generated catalogue of ~200 integration tools (vectorshift.agent.tools), and Session objects with streamed events.

​
What’s next

Reference

Every public method, grouped by topic.

Examples

Runnable examples to copy-paste and adapt.

Session

Multi-turn streaming on top of an Agent.