SDK reference

The AgentRoute SDK is one import away. Everything you need to build, run, and reason about an agent lives under the top-level agentroute package, so you can pull symbols straight from the namespace:

from agentroute import Agent, tool, Context, Memory, Result

This is the reference for the public API as of __version__ == "1.1.0". Each page below documents one module of the SDK with exact signatures, defaults, return types, and runnable examples. If you are new here, start with the Agent page, or read the concepts guide for the bigger picture before diving into signatures.

Reference pages

Agent

The agentic loop. Construct an agent, register tools, run it, and get a Result back.

Tools

The @tool decorator and the Tool dataclass. Turn any Python function into a model-callable tool.

Context & Usage

The per-run Context injected into tools, and the Usage accumulator for tokens and cost.

Memory

In-RAM Memory and persistent MemorySQLite, plus the MemoryProto protocol they implement.

History

Conversation compaction policies: sliding window, truncation, and LLM summarization.

Models

The Model protocol, OpenAI-compatible message types, and resolve_model for any provider string.

Results

The Result your run returns, plus Event for streaming and ResultDeploy for deployment.

Exceptions

The error hierarchy — ErrorAgent, ErrorMaxTurns, ErrorBudget — and the Retry control signal.

Every export, by page

The agentroute package exports 24 public symbols. Here is each one and the reference page that documents it.

Symbol	Kind	Documented on
`Agent`	class	Agent
`tool`	function	Tools
`Tool`	dataclass	Tools
`Context`	dataclass	Context & Usage
`Usage`	dataclass	Context & Usage
`Memory`	class	Memory
`MemorySQLite`	class	Memory
`History`	protocol	History
`HistorySlidingWindow`	class	History
`HistoryTruncate`	class	History
`HistorySummarize`	class	History
`Model`	protocol	Models
`resolve_model`	function	Models
`Message`	dataclass	Models
`LLMRequest`	dataclass	Models
`LLMResponse`	dataclass	Models
`Result`	dataclass	Results
`Event`	dataclass	Results
`ResultDeploy`	dataclass	Results
`ErrorAgent`	exception	Exceptions
`ErrorMaxTurns`	exception	Exceptions
`ErrorBudget`	exception	Exceptions
`Retry`	exception	Exceptions

One import to rule them all

Every symbol in the table is re-exported from the package root, so from agentroute import Agent, MemorySQLite, ErrorMaxTurns always works. You never need to reach into submodules like agentroute.memory.history.

A complete example

Here is a minimal agent that wires together symbols from several pages — a tool, structured output, and a typed Result:

from agentroute import Agent
from pydantic import BaseModel
 
 
class Weather(BaseModel):
    city: str
    summary: str
 
 
agent = Agent(name="forecaster", model="claude-sonnet-4", output=Weather)
 
 
@agent.tool
def get_temp(city: str) -> int:
    """Return the current temperature in Celsius for a city."""
    return 21
 
 
result = agent.run("What's the weather in Berlin?")
print(result.output.summary)  # -> a Weather instance
print(result.usage.total_cost_usd)  # -> accumulated cost for the run

From here, pick a page above to see the full surface of each module.

Source: agentroute/init.py