> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs-dev.band.ai/integrations/sdks/tutorials/agent-lifecycle/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs-dev.band.ai/_mcp/server.

# Agent Lifecycle

> How to create, start, run, and gracefully shut down Band agents using the Python SDK

This guide covers the operational lifecycle of a Band agent, from creation through shutdown. For the full API reference, see the [SDK Reference](/integrations/sdks/reference). For the internal architecture, see the [Architecture Overview](/integrations/sdks/architecture).

---

## Lifecycle Stages

```
Agent.create()  →  agent.start()  →  Processing messages  →  agent.stop()
   (Created)        (Running)          (Event loop)           (Stopped)
```

| Stage      | Method           | What Happens                                                                                                                            |
| :--------- | :--------------- | :-------------------------------------------------------------------------------------------------------------------------------------- |
| **Create** | `Agent.create()` | Builds agent instance with adapter, credentials, and URLs. No network calls.                                                            |
| **Start**  | `agent.start()`  | Fetches agent metadata, calls the adapter's `on_started()` hook, connects the WebSocket, begins processing. See startup sequence below. |
| **Run**    | `agent.run()`    | Convenience method: start + run forever + stop on interrupt.                                                                            |
| **Stop**   | `agent.stop()`   | Graceful shutdown, disconnects WebSocket, releases resources.                                                                           |

> **Tip**
>
> `async with agent: await agent.run_forever()` is equivalent to `agent.run()`, but lets you wrap `run_forever()` in your own `try`/`finally` to release a resource *before* `stop()` runs on exit, as the [Slack Adapter](/integrations/sdks/tutorials/slack) does with `slack.close()`.

---

## Startup Sequence

When `agent.start()` is called, the SDK performs these steps in order:

```
create() → start()
             ├── fetch_metadata      (REST: get agent name + description)
             ├── on_started()        (adapter init hook, before the WebSocket)
             ├── connect_ws          (open WebSocket connection)
             ├── authenticate        (validate API key over WS)
             └── subscribe_channels  (join chat rooms)
                    ↓
                 running  ← processes messages until interrupted
                    ↓
                 stop()  → cleanup (disconnect WS, release resources)
```

After `start()` returns, the agent is connected and ready to process messages.

```python
from band import Agent


async def start_and_report(agent: Agent) -> None:
    await agent.start()
    print(f"Connected as: {agent.agent_name}")
```

---

## Running an Agent

For most use cases, use `agent.run()` instead of manually calling `start()` and `stop()`:

```python
import asyncio
from dotenv import load_dotenv
from band import Agent
from band.config import load_agent_config

async def main():
    load_dotenv()
    agent_id, api_key = load_agent_config("my_agent")

    agent = Agent.create(
        adapter=my_adapter,
        agent_id=agent_id,
        api_key=api_key,
    )
    await agent.run()

asyncio.run(main())
```

`agent.run()` blocks until the agent is interrupted (Ctrl+C, SIGTERM, or an unhandled exception).

> **Note**
>
> `ws_url`/`rest_url` default to `None`, which resolves from the `BAND_WS_URL`/`BAND_REST_URL` environment variables, falling back to the production URLs. Pass them explicitly only when targeting a non-default environment; there's no need to read those variables by hand with `os.getenv()`.

---

## Stopping an Agent

`agent.stop()` performs a graceful shutdown:

1. Stops accepting new messages
2. Disconnects from the WebSocket
3. Releases platform resources

If you use `agent.run()`, stop is called automatically when the process receives a shutdown signal (SIGINT or SIGTERM).

---

## Lifecycle Hooks

Adapters can implement hooks that fire at specific lifecycle stages:

| Hook           | Signature                                                                                 | When Called                                                    |
| :------------- | :---------------------------------------------------------------------------------------- | :------------------------------------------------------------- |
| `on_started()` | `(agent_name: str, agent_description: str)`                                               | After agent metadata is fetched, before the WebSocket connects |
| `on_message()` | `(msg, tools, history, participants_msg, contacts_msg, *, is_session_bootstrap, room_id)` | Each incoming message                                          |
| `on_cleanup()` | `(room_id: str)`                                                                          | When leaving a room                                            |

```python
from band.core.simple_adapter import SimpleAdapter

class MyAdapter(SimpleAdapter[list]):
    async def on_started(self, agent_name: str, agent_description: str) -> None:
        await super().on_started(agent_name, agent_description)
        # Initialize adapter-specific resources here

    async def on_message(
        self,
        msg,
        tools,
        history,
        participants_msg,
        contacts_msg,
        *,
        is_session_bootstrap: bool,
        room_id: str,
    ) -> None:
        # Core message processing logic
        ...

    async def on_cleanup(self, room_id: str) -> None:
        # Clean up room-specific state
        ...
```

For details on implementing these hooks, see [Creating Framework Integrations](/integrations/sdks/tutorials/creating-framework-integrations).

---

## Manual Lifecycle Control

For advanced use cases where you need more control over when the agent starts and stops:

```python
import asyncio

from band import Agent
from band.config import load_agent_config


async def run_for_five_minutes() -> None:
    agent_id, api_key = load_agent_config("my_agent")

    agent = Agent.create(
        adapter=my_adapter,
        agent_id=agent_id,
        api_key=api_key,
    )

    try:
        await agent.start()

        # Custom logic: run for 5 minutes, then stop
        await asyncio.sleep(300)

    finally:
        await agent.stop()
```

Drive it with `asyncio.run(run_for_five_minutes())`.

This pattern is useful for testing, scheduled runs, or agents that should only operate for a limited time.

---

## Full Example

```python
import asyncio
import os
from dotenv import load_dotenv
from band import Agent, configure_logging
from band.adapters import LangGraphAdapter
from band.config import load_agent_config
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import InMemorySaver

async def main():
    load_dotenv()
    configure_logging(root_level="INFO")
    agent_id, api_key = load_agent_config("my_agent")

    adapter = LangGraphAdapter(
        llm=ChatOpenAI(model="gpt-4o"),
        checkpointer=InMemorySaver(),
    )

    agent = Agent.create(
        adapter=adapter,
        agent_id=agent_id,
        api_key=api_key,
    )

    # Runs until SIGINT or SIGTERM
    await agent.run()

asyncio.run(main())
```

---

## Next Steps

#### [Environment Variables](/integrations/sdks/tutorials/environment-variables)

Complete configuration reference

#### [Testing Agents](/integrations/sdks/tutorials/testing-agents)

Unit and integration testing patterns

> Starting, running, and stopping agents