Runtime

The Runtime is the global orchestrator that manages all agents, sessions, and execution across Agent Kernel. You can skip this section if you are not planning to contribute to Aent Kernel.

Overview

What is the Runtime?

The Runtime:

• Maintains global agent registry
• Manages sessions across requests
• Provides centralized configuration
• Coordinates execution across modes (CLI, API, Lambda)
• Enables service integration (MCP, A2A)

Singleton Pattern

The Runtime uses a singleton pattern - there's only one instance:

from agentkernel.core import Runtime

# Always returns the same instance
runtime1 = Runtime.current()
runtime2 = Runtime.current()

assert runtime1 is runtime2  # True

Alternatively you can use it as the context manager for advanced use cases (see below)

Accessing Agents

Get Agent by Name

from agentkernel.core import Runtime

runtime = Runtime.current()
agent = runtime.agents().get("assistant")

Get All Agents

runtime = Runtime.current()
all_agents = runtime.agents()

for name, agent in all_agents.items():
    print(f"Agent: {name}")

Check Agent Existence

runtime = Runtime.current()

if "assistant" in runtime.agents():
    agent = runtime.agents()["assistant"]
else:
    print("Agent not found")

Session Management

The Runtime manages sessions through a SessionStore:

from agentkernel.core import Runtime

runtime = Runtime.current()

# Get existing session or create new one
session = runtime.sessions().get("user-123")
if session is None:
    session = runtime.sessions().new("user-123")

# Session is automatically persisted based on configuration

Note: For detailed information about session management, storage backends, and configuration, see the Session Management documentation.

Configuration

Configuration is accessed through the AKConfig singleton:

from agentkernel.core.config import AKConfig

config = AKConfig.get()

print(config.log_level)
print(config.session.type)  # 'in_memory', 'redis', or 'dynamodb'

Execution Modes

The Runtime supports multiple execution modes:

CLI Mode

from agentkernel.cli import CLI

# CLI uses Runtime to discover and execute agents
CLI.main()

REST API Mode

from agentkernel.api import RESTAPI

# API server uses Runtime to route requests
RESTAPI.run()

AWS Lambda Mode

from agentkernel.aws import Lambda

# Lambda handler uses Runtime to process events
handler = Lambda.handler

MCP Server Mode

from agentkernel.mcp import MCP

# MCP server exposes agents via Runtime
server = MCP.get()  

Runtime Lifecycle

Advanced Usage

Custom Runtime Context

For advanced use cases, you can create a custom runtime instance with specific configuration and use it as a context manager:

from agentkernel.core import Runtime
from agentkernel.core.builder import SessionStoreBuilder

# Create a custom runtime instance
custom_runtime = Runtime(SessionStoreBuilder.build())

# Use it within a context
with custom_runtime:
    # Within this context, Runtime.current() returns custom_runtime
    agent = MyAgent("custom-agent")
    custom_runtime.register(agent)
    session = custom_runtime.sessions().new("session-1")
    
    # Execute agent with this runtime
    from agentkernel.core.model import AgentRequestText
    requests = [AgentRequestText(text="Hello")]
    result = await custom_runtime.run(agent, session, requests)

Accessing the Current Runtime

The Runtime.current() static method returns the currently active runtime instance. If called within a runtime context manager, it returns that specific runtime; otherwise, it returns the global singleton:

from agentkernel.core import Runtime
from agentkernel.core.runtime import GlobalRuntime

# Outside any context, returns GlobalRuntime
current = Runtime.current()
assert current == GlobalRuntime.instance()

# Inside a context, returns that runtime
with custom_runtime:
    current = Runtime.current()
    assert current == custom_runtime

This pattern is particularly useful when:

• Writing framework integrations that need access to the active runtime
• Building utilities that work with whichever runtime is currently active
• Testing with isolated runtime instances

Custom Agent Registration

Manually register agents (advanced):

from agentkernel.core import Runtime

runtime = Runtime.current()

# Manually register an agent
runtime.register(custom_agent)

Integration Points

MCP Integration

# when MCP server is enabled
# AK_MCP_ENABLED=true

A2A Integration

# for all registered agents
# AK_A2A_ENABLED=true

REST API Integration

# GET /api/v1/agents - list all agents
# POST /api/v1/chat - execute agent

Best Practices

Single Runtime Instance

Always use Runtime.current() for generic usecases:

# Correct
runtime = Runtime.current()

# Don't instantiate directly unless you have a specific need
# runtime = Runtime(sessions)  # Only for advanced use cases

Configuration Before Execution

Set environment variables before importing:

import os
os.environ["AK_SESSION_STORAGE"] = "redis"
os.environ["AK_REDIS_URL"] = "redis://localhost:6379"

# Now import and use
from agentkernel.core import Runtime
runtime = Runtime.current()

Summary

• Runtime is the global orchestrator
• Maintains agent registry
• Manages sessions
• Provides centralized configuration
• Supports multiple execution modes
• Use Runtime.current() to access the active runtime instance
• Can be used as a context manager for isolated runtime contexts
• Thread-safe runtime state management with internal locking

Next Steps

• Session Management - Detailed session configuration and lifecycle
• Deployment Overview
• REST API
• Configuration