Agent

Agent

Agent is the canonical runtime entry point in Agiwo. It owns one model, one config object, and one assembled tool surface.

Public execution primitives

• run(...) for one-shot request and response
• run_stream(...) for streaming consumption
• start(...) for an explicit execution handle

All three use the same underlying runtime pipeline.

Construction

from agiwo.agent import Agent, AgentConfig
from agiwo.llm import OpenAIModel

agent = Agent(
    AgentConfig(
        name="assistant",
        description="Helpful SDK assistant",
        system_prompt="Answer concisely and use tools when needed.",
    ),
    model=OpenAIModel(name="gpt-5.4"),
)

Important config semantics

• allowed_tools=None means builtin functional tools plus all extra functional tools
• allowed_tools=[] means no functional tools
• allowed_skills=None means all discovered skills remain eligible
• allowed_skills=[] disables skills

allowed_skills must already be expanded to explicit skill names before entering runtime.

Stable identity

Pass a stable id when the same logical agent is reconstructed across requests, especially in Console-backed or persistent scheduler flows. That keeps history and runtime state aligned.

Tool assembly

An agent can expose:

• builtin functional tools
• extra tools passed through tools=[...]
• skills, when enabled
• scheduler runtime tools, when running under Scheduler

The scheduler tools are system-owned and are not manually registered.

Lifecycle

Agents hold live model and storage resources. Call await agent.close() when you create agents directly and no longer need them.