Agents & Assistants
This guide focuses on how to implement Agents & Assistants using Bolt. For general information about the feature, please refer to the API documentation.
To get started, enable the Agents & Assistants feature on the app configuration page. Add assistant:write, chat:write, and im:history to the bot scopes on the OAuth & Permissions page. Make sure to subscribe to assistant_thread_started, assistant_thread_context_changed, and message.im events on the Event Subscriptions page.
Please note that this feature requires a paid plan. If you don't have a paid workspace for development, you can join the Developer Program and provision a sandbox with access to all Slack features for free.
To handle assistant thread interactions with humans, although you can implement your agents using app.event(...) listeners for assistant_thread_started, assistant_thread_context_changed, and message events, Bolt offers a simpler approach. You just need to create an Assistant instance, attach the needed event handlers to it, and then add the assistant to your App instance.
assistant = Assistant()
# This listener is invoked when a human user opened an assistant thread
@assistant.thread_started
def start_assistant_thread(say: Say, set_suggested_prompts: SetSuggestedPrompts):
# Send the first reply to the human who started chat with your app's assistant bot
say(":wave: Hi, how can I help you today?")
# Setting suggested prompts is optional
set_suggested_prompts(
prompts=[
# If the suggested prompt is long, you can use {"title": "short one to display", "message": "full prompt"} instead
"What does SLACK stand for?",
"When Slack was released?",
],
)
# This listener is invoked when the human user sends a reply in the assistant thread
@assistant.user_message
def respond_in_assistant_thread(
payload: dict,
logger: logging.Logger,
context: BoltContext,
set_status: SetStatus,
client: WebClient,
say: Say,
):
try:
# Tell the human user the assistant bot acknowledges the request and is working on it
set_status("is typing...")
# Collect the conversation history with this user
replies_in_thread = client.conversations_replies(
channel=context.channel_id,
ts=context.thread_ts,
oldest=context.thread_ts,
limit=10,
)
messages_in_thread: List[Dict[str, str]] = []
for message in replies_in_thread["messages"]:
role = "user" if message.get("bot_id") is None else "assistant"
messages_in_thread.append({"role": role, "content": message["text"]})
# Pass the latest prompt and chat history to the LLM (call_llm is your own code)
returned_message = call_llm(messages_in_thread)
# Post the result in the assistant thread
say(text=returned_message)
except Exception as e:
logger.exception(f"Failed to respond to an inquiry: {e}")
# Don't forget sending a message telling the error
# Without this, the status 'is typing...' won't be cleared, therefore the end-user is unable to continue the chat
say(f":warning: Sorry, something went wrong during processing your request (error: {e})")
# Enable this assistant middleware in your Bolt app
app.use(assistant)
Refer to the module document to learn the available listener arguments.
When a user opens an Assistant thread while in a channel, the channel information is stored as the thread's AssistantThreadContext data. You can access this information by using the get_thread_context utility. The reason Bolt provides this utility is that the most recent thread context information is not included in the subsequent user message event payload data. Therefore, an app must store the context data when it is changed so that the app can refer to the data in message event listeners.
When the user switches channels, the assistant_thread_context_changed event will be sent to your app. If you use the built-in Assistant middleware without any custom configuration (like the above code snippet does), the updated context data is automatically saved as message metadata of the first reply from the assistant bot.
As long as you use the built-in approach, you don't need to store the context data within a datastore. The downside of this default behavior is the overhead of additional calls to the Slack API. These calls include those to conversations.history which are used to look up the stored message metadata that contains the thread context (via get_thread_context).
To store context elsewhere, pass a custom AssistantThreadContextStore implementation to the Assistant constructor. We provide FileAssistantThreadContextStore, which is a reference implementation that uses the local file system:
# You can use your own thread_context_store if you want
from slack_bolt import FileAssistantThreadContextStore
assistant = Assistant(thread_context_store=FileAssistantThreadContextStore())
Since this reference implementation relies on local files, it's not advised for use in production. For production apps, we recommend creating a class that inherits AssistantThreadContextStore.
Block Kit interactions in the assistant thread
For advanced use cases, Block Kit buttons may be used instead of suggested prompts, as well as the sending of messages with structured metadata to trigger subsequent interactions with the user.
For example, an app can display a button like "Summarize the referring channel" in the initial reply. When the user clicks the button and submits detailed information (such as the number of messages, days to check, the purpose of the summary, etc.), the app can handle that information and post a message that describes the request with structured metadata.
By default, apps can't respond to their own bot messages (Bolt prevents infinite loops by default). However, if you pass ignoring_self_assistant_message_events_enabled=False to the App constructor and add a bot_message listener to your Assistant middleware, your app can continue processing the request as shown below:
app = App(
token=os.environ["SLACK_BOT_TOKEN"],
# This must be set to handle bot message events
ignoring_self_assistant_message_events_enabled=False,
)
assistant = Assistant()
# Refer to https://tools.slack.dev/bolt-python/api-docs/slack_bolt/kwargs_injection/args.html to learn available listener arguments
@assistant.thread_started
def start_assistant_thread(say: Say):
say(
text=":wave: Hi, how can I help you today?",
blocks=[
{
"type": "section",
"text": {"type": "mrkdwn", "text": ":wave: Hi, how can I help you today?"},
},
{
"type": "actions",
"elements": [
# You can have multiple buttons here
{
"type": "button",
"action_id": "assistant-generate-random-numbers",
"text": {"type": "plain_text", "text": "Generate random numbers"},
"value": "clicked",
},
],
},
],
)
# This listener is invoked when the above button is clicked
@app.action("assistant-generate-random-numbers")
def configure_random_number_generation(ack: Ack, client: WebClient, body: dict):
ack()
client.views_open(
trigger_id=body["trigger_id"],
view={
"type": "modal",
"callback_id": "configure_assistant_summarize_channel",
"title": {"type": "plain_text", "text": "My Assistant"},
"submit": {"type": "plain_text", "text": "Submit"},
"close": {"type": "plain_text", "text": "Cancel"},
# Relay the assistant thread information to app.view listener
"private_metadata": json.dumps(
{
"channel_id": body["channel"]["id"],
"thread_ts": body["message"]["thread_ts"],
}
),
"blocks": [
{
"type": "input",
"block_id": "num",
"label": {"type": "plain_text", "text": "# of outputs"},
# You can have this kind of predefined input from a user instead of parsing human text
"element": {
"type": "static_select",
"action_id": "input",
"placeholder": {"type": "plain_text", "text": "How many numbers do you need?"},
"options": [
{"text": {"type": "plain_text", "text": "5"}, "value": "5"},
{"text": {"type": "plain_text", "text": "10"}, "value": "10"},
{"text": {"type": "plain_text", "text": "20"}, "value": "20"},
],
"initial_option": {"text": {"type": "plain_text", "text": "5"}, "value": "5"},
},
}
],
},
)
# This listener is invoked when the above modal is submitted
@app.view("configure_assistant_summarize_channel")
def receive_random_number_generation_details(ack: Ack, client: WebClient, payload: dict):
ack()
num = payload["state"]["values"]["num"]["input"]["selected_option"]["value"]
thread = json.loads(payload["private_metadata"])
# Post a bot message with structured input data
# The following assistant.bot_message will continue processing
# If you prefer processing this request within this listener, it also works!
# If you don't need bot_message listener, no need to set ignoring_self_assistant_message_events_enabled=False
client.chat_postMessage(
channel=thread["channel_id"],
thread_ts=thread["thread_ts"],
text=f"OK, you need {num} numbers. I will generate it shortly!",
metadata={
"event_type": "assistant-generate-random-numbers",
"event_payload": {"num": int(num)},
},
)
# This listener is invoked whenever your app's bot user posts a message
@assistant.bot_message
def respond_to_bot_messages(logger: logging.Logger, set_status: SetStatus, say: Say, payload: dict):
try:
if payload.get("metadata", {}).get("event_type") == "assistant-generate-random-numbers":
# Handle the above random-number-generation request
set_status("is generating an array of random numbers...")
time.sleep(1)
nums: Set[str] = set()
num = payload["metadata"]["event_payload"]["num"]
while len(nums) < num:
nums.add(str(random.randint(1, 100)))
say(f"Here you are: {', '.join(nums)}")
else:
# nothing to do for this bot message
# If you want to add more patterns here, be careful not to cause infinite loop messaging
pass
except Exception as e:
logger.exception(f"Failed to respond to an inquiry: {e}")
# This listener is invoked when the human user posts a reply
@assistant.user_message
def respond_to_user_messages(logger: logging.Logger, set_status: SetStatus, say: Say):
try:
set_status("is typing...")
say("Please use the buttons in the first reply instead :bow:")
except Exception as e:
logger.exception(f"Failed to respond to an inquiry: {e}")
say(f":warning: Sorry, something went wrong during processing your request (error: {e})")
# Enable this assistant middleware in your Bolt app
app.use(assistant)
Lastly, if you want to check full working example app, you can check our sample repository on GitHub.