agentica-sdk（agentica SDK 搜索）

Agentica SDK Reference (v0.3.1)

Build AI agents in Python using the Agentica framework. Agents can implement functions, maintain state, use tools, and coordinate with each other.

When to Use

Use this skill when:

• Building new Python agents
• Adding agentic capabilities to existing code
• Integrating MCP tools with agents
• Implementing multi-agent orchestration
• Debugging agent behavior

Quick Start

Agentic Function (simplest)

from agentica import agentic

@agentic()
async def add(a: int, b: int) -> int:
    """Returns the sum of a and b"""
    ...

result = await add(1, 2)  # Agent computes: 3

Spawned Agent (more control)

from agentica import spawn

agent = await spawn(premise="You are a truth-teller.")
result: bool = await agent.call(bool, "The Earth is flat")
# Returns: False

Core Patterns

Return Types

# String (default)
result = await agent.call("What is 2+2?")

# Typed output
result: int = await agent.call(int, "What is 2+2?")
result: dict[str, int] = await agent.call(dict[str, int], "Count items")

# Side-effects only
await agent.call(None, "Send message to John")

Premise vs System Prompt

# Premise: adds to default system prompt
agent = await spawn(premise="You are a math expert.")

# System: full control (replaces default)
agent = await spawn(system="You are a JSON-only responder.")

Passing Tools (Scope)

from agentica import agentic, spawn

# In decorator
@agentic(scope={'web_search': web_search_fn})
async def researcher(query: str) -> str:
    """Research a topic."""
    ...

# In spawn
agent = await spawn(
    premise="Data analyzer",
    scope={"analyze": custom_analyzer}
)

# Per-call scope
result = await agent.call(
    dict[str, int],
    "Analyze the dataset",
    dataset=data,           # Available as 'dataset'
    analyzer=custom_fn      # Available as 'analyzer'
)

SDK Integration Pattern

from slack_sdk import WebClient

slack = WebClient(token=SLACK_TOKEN)

# Extract specific methods
@agentic(scope={
    'list_users': slack.users_list,
    'send_message': slack.chat_postMessage
})
async def team_notifier(message: str) -> None:
    """Send team notifications."""
    ...

Agent Instantiation

spawn() - Async (most cases)

agent = await spawn(premise="Helpful assistant")

Agent() - Sync (for __init__)

from agentica.agent import Agent

class CustomAgent:
    def __init__(self):
        # Synchronous - use Agent() not spawn()
        self._brain = Agent(
            premise="Specialized assistant",
            scope={"tool": some_tool}
        )

    async def run(self, task: str) -> str:
        return await self._brain(str, task)

Model Selection

# In spawn
agent = await spawn(
    premise="Fast responses",
    model="openai:gpt-5"  # Default: openai:gpt-4.1
)

# In decorator
@agentic(model="anthropic:claude-sonnet-4.5")
async def analyze(text: str) -> dict:
    """Analyze text."""
    ...

Available models:

• openai:gpt-3.5-turbo, openai:gpt-4o, openai:gpt-4.1, openai:gpt-5
• anthropic:claude-sonnet-4, anthropic:claude-opus-4.1
• anthropic:claude-sonnet-4.5, anthropic:claude-opus-4.5
• Any OpenRouter slug (e.g., google/gemini-2.5-flash)

Persistence (Stateful Agents)

@agentic(persist=True)
async def chatbot(message: str) -> str:
    """Remembers conversation history."""
    ...

await chatbot("My name is Alice")
await chatbot("What's my name?")  # Knows: Alice

For spawn() agents, state is automatic across calls to the same instance.

Token Limits

from agentica import spawn, MaxTokens

# Simple limit
agent = await spawn(
    premise="Brief responses",
    max_tokens=500
)

# Fine-grained control
agent = await spawn(
    premise="Controlled output",
    max_tokens=MaxTokens(
        per_invocation=5000,  # Total across all rounds
        per_round=1000,       # Per inference round
        rounds=5              # Max inference rounds
    )
)

Token Usage Tracking

from agentica import spawn, last_usage, total_usage

agent = await spawn(premise="You are helpful.")
await agent.call(str, "Hello!")

# Agent method
usage = agent.last_usage()
print(f"Last: {usage.input_tokens} in, {usage.output_tokens} out")

usage = agent.total_usage()
print(f"Total: {usage.total_tokens} processed")

# For @agentic functions
@agentic()
async def my_fn(x: str) -> str: ...

await my_fn("test")
print(last_usage(my_fn))
print(total_usage(my_fn))

Streaming

from agentica import spawn
from agentica.logging.loggers import StreamLogger
import asyncio

agent = await spawn(premise="You are helpful.")

stream = StreamLogger()
with stream:
    result = asyncio.create_task(
        agent.call(bool, "Is Paris the capital of France?")
    )

# Consume stream FIRST for live output
async for chunk in stream:
    print(chunk.content, end="", flush=True)
# chunk.role is 'user', 'agent', or 'system'

# Then await result
final = await result

MCP Integration

from agentica import spawn, agentic

# Via config file
agent = await spawn(
    premise="Tool-using agent",
    mcp="path/to/mcp_config.json"
)

@agentic(mcp="path/to/mcp_config.json")
async def tool_user(query: str) -> str:
    """Uses MCP tools."""
    ...

mcp_config.json format:

{
  "mcpServers": {
    "tavily-remote-mcp": {
      "command": "npx -y mcp-remote https://mcp.tavily.com/mcp/?tavilyApiKey=<key>",
      "env": {}
    }
  }
}

Logging

Default Behavior

• Prints to stdout with colors
• Writes to ./logs/agent-<id>.log

Contextual Logging

from agentica.logging.loggers import FileLogger, PrintLogger
from agentica.logging.agent_logger import NoLogging

# File only
with FileLogger():
    agent = await spawn(premise="Debug agent")
    await agent.call(int, "Calculate")

# Silent
with NoLogging():
    agent = await spawn(premise="Silent agent")

Per-Agent Logging

# Listeners are in agent_listener submodule (NOT exported from agentica.logging)
from agentica.logging.agent_listener import (
    PrintOnlyListener,  # Console output only
    FileOnlyListener,   # File logging only
    StandardListener,   # Both console + file (default)
    NoopListener,       # Silent - no logging
)

agent = await spawn(
    premise="Custom logging",
    listener=PrintOnlyListener
)

# Silent agent
agent = await spawn(
    premise="Silent agent",
    listener=NoopListener
)

Global Config

from agentica.logging.agent_listener import (
    set_default_agent_listener,
    get_default_agent_listener,
    PrintOnlyListener,
)

set_default_agent_listener(PrintOnlyListener)
set_default_agent_listener(None)  # Disable all

Error Handling

from agentica.errors import (
    AgenticaError,           # Base for all SDK errors
    RateLimitError,          # Rate limiting
    InferenceError,          # HTTP errors from inference
    MaxTokensError,          # Token limit exceeded
    MaxRoundsError,          # Max inference rounds exceeded
    ContentFilteringError,   # Content filtered
    APIConnectionError,      # Network issues
    APITimeoutError,         # Request timeout
    InsufficientCreditsError,# Out of credits
    OverloadedError,         # Server overloaded
    ServerError,             # Generic server error
)

try:
    result = await agent.call(str, "Do something")
except RateLimitError:
    await asyncio.sleep(60)
    result = await agent.call(str, "Do something")
except MaxTokensError:
    # Reduce scope or increase limits
    pass
except ContentFilteringError:
    # Content was filtered
    pass
except InferenceError as e:
    logger.error(f"Inference failed: {e}")
except AgenticaError as e:
    logger.error(f"SDK error: {e}")

Custom Exceptions

class DataValidationError(Exception):
    """Invalid input data."""
    pass

@agentic(DataValidationError)  # Pass exception type
async def analyze(data: str) -> dict:
    """
    Analyze data.

    Raises:
        DataValidationError: If data is malformed
    """
    ...

try:
    result = await analyze(raw_data)
except DataValidationError as e:
    logger.warning(f"Invalid: {e}")

Multi-Agent Patterns

Custom Agent Class

from agentica.agent import Agent

class ResearchAgent:
    def __init__(self, web_search_fn):
        self._brain = Agent(
            premise="Research assistant.",
            scope={"web_search": web_search_fn}
        )

    async def research(self, topic: str) -> str:
        return await self._brain(str, f"Research: {topic}")

    async def summarize(self, text: str) -> str:
        return await self._brain(str, f"Summarize: {text}")

Agent Orchestration

class LeadResearcher:
    def __init__(self):
        self._brain = Agent(
            premise="Coordinate research across subagents.",
            scope={"SubAgent": ResearchAgent}
        )

    async def __call__(self, query: str) -> str:
        return await self._brain(str, query)

lead = LeadResearcher()
report = await lead("Research AI agent frameworks 2025")

Tracing & Debugging

OpenTelemetry Tracing

from agentica import initialize_tracing

# Initialize tracing (returns TracerProvider)
tracer = initialize_tracing(
    service_name="my-agent-app",
    environment="development",  # Optional
    tempo_endpoint="http://localhost:4317",  # Optional: Grafana Tempo
    organization_id="my-org",  # Optional
    log_level="INFO",  # DEBUG, INFO, WARNING, ERROR
    instrument_httpx=False,  # Optional: trace HTTP calls
)

SDK Debug Logging

from agentica import enable_sdk_logging

# Enable internal SDK logs (for debugging the SDK itself)
disable_fn = enable_sdk_logging(log_tags="1")

# ... run agents ...

disable_fn()  # Disable when done

Top-Level Exports

# Main imports from agentica
from agentica import (
    # Core
    Agent,              # Synchronous agent class
    agentic,            # @agentic decorator
    spawn,              # Async agent creation

    # Configuration
    ModelStrings,       # Model string type hints
    AgenticFunction,    # Agentic function type

    # Token tracking
    last_usage,         # Get last call's token usage
    total_usage,        # Get cumulative token usage

    # Tracing/Logging
    initialize_tracing, # OpenTelemetry setup
    enable_sdk_logging, # SDK debug logs

    # Version
    __version__,        # "0.3.1"
)

Checklist

Before using Agentica:

• Functions with @agentic() MUST be async
• spawn() returns awaitable - use await spawn(...)
• agent.call() is awaitable - use await agent.call(...)
• First arg to call() is return type, second is prompt string
• Use persist=True for conversation memory in @agentic
• Use Agent() (not spawn()) in synchronous __init__
• Document exceptions in docstrings for agent to raise them
• Import listeners from agentica.logging.agent_listener (NOT agentica.logging)