Core concepts

Agents

An agent proposes actions; the runtime governs and executes them. Antraft ships a ReAct agent and prebuilt patterns, and any object that proposes actions can be governed.

ReActAgent

The default agent runs an LLM tool-calling loop, emitting neutral Action andThought objects so every tool call flows through the Guard.

agent.py
from antraft import Antraft, ToolRegistry, tool_from_function
from antraft.models import create_model
 
runtime = (
    Antraft.agent(
        create_model("openai:gpt-4o"),
        ToolRegistry([tool_from_function(my_tool)]),
        task="...",
        system_prompt="You are concise.",
        max_turns=12,
        parallel_tools=True,   # one turn's tool calls run concurrently
    )
    .allow(["my_tool"])
    .build()
)

Structured output

Pass a pydantic model or JSON Schema; the final answer is validated into it.

structured.py
from pydantic import BaseModel
 
class Summary(BaseModel):
    title: str
    bullets: list[str]
 
runtime = Antraft.agent(model, tools, task="...", output_schema=Summary).build()
await runtime.run()
runtime.agent.structured        # -> Summary(...)  (or .structured_error)

Prebuilt patterns

patterns.py
from antraft.agents import reflection_agent, plan_execute_agent
 
# self-critique then improve, before finishing
agent = reflection_agent(model, tools, task="...", reflect_rounds=1)
 
# plan explicitly, then execute the plan
agent = plan_execute_agent(model, tools, task="...")

Govern any agent

Implement next_action() and observe(result) and Antraft will govern it — no need to use ReActAgent.

custom.py
from antraft.core.agent import BaseAgent
from antraft.core.action import Action
 
class MyAgent(BaseAgent):
    id = "my-agent"
    def next_action(self):
        return Action(name="search", params={"q": "antraft"})  # or None to finish
    def observe(self, result):
        ...
 
await Antraft.guard(MyAgent(), {"search": do_search}).allow(["search"]).run()
Tip
Multi-agent workflows compose with antraft.graph.Graph — nodes can run governed agents, edges route on shared state.
Next →
Tools