Tips for chatting with the Cortex Agentic Assistant
After choosing an agent, type a request using natural language. Be as clear and specific as possible. Submit your request by pressing Enter or clicking the submit arrow. Some agents provide relevant chat conversation starters under the chat prompt.
During a conversation, when an agent is formulating a plan or executing steps, clicking the agent will show which actions it is using. You can scroll between the actions or close the panel.
Prompt examples
Using chat prompt conversation starters in the Agentic Assistant simplifies and speeds up your interactions by providing pre-defined, common queries that guide you to relevant actions and information.
For example, a SOC analyst may see the following conversation starters under the chat prompt:
What are the top issues I should prioritize today?
Show me all issues with an overdue SLA
Which automations are waiting for my input?
Clean up all expired indicators.
Create a visual representation of the top 10 targeted assets over the last 7 days.
Best practices for prompting
Be clear and specific
Clearly state your objective and provide the necessary context. Instead of "Fix the issue," try "Investigate issue 1234 and isolate any affected hosts on which malware has been identified." Specify exact values, IDs, and relevant details.
Break down complex tasks
For multi-step processes, break your request into smaller steps within a prompt. This allows the agent to focus, validate each step, and helps guide the flow.
Include key information
If not available from the current case context, always include relevant incident IDs, indicator values (such as IP addresses and file hashes), or entity names directly in your prompt. The more precise the initial information, the better the agent can leverage its actions and context.
Specify a desired output or action:
If you need a particular type of output (for example, "Summarize the findings," "List all affected assets") or a specific action (for example, "Isolate host X," "Block IP Y on firewall"), explicitly state it.
Considerations for Slack interactions with Agentic Assistant agents
When interacting with the Cortex Agentic Assistant directly within Slack, keep the following in mind to maintain session integrity and secure access:
Chat context
When an agent is tagged, the system automatically pulls in the last five messages in the thread (or up to the last bot interaction) so the agent understands the conversation's history.
Single player model
The first user to tag the bot becomes the initiator, and only this user can issue commands. If another user tries to send a prompt, they receive an Access Denied message.
The reset command
To hand off a session to another user or start fresh, any user in the Slack thread can type
@Cortex Assistant reset. This ends the session and allows a new initiator to take over.Approving sensitive actions (hard locks)
Sensitive actions require approval. In Slack, this triggers a hard lock where the agent refuses text input and displays Approve (green) and Deny (red) buttons, and only the initiator can click these buttons. If the initiator does not respond within two weeks, the request is automatically denied and the chat will close.
Providing feedback
After a final result or remediation is executed, you can provide feedback directly in Slack using thumbs up or thumbs down.
Additional considerations
Small tables (less than 5 rows) are rendered as Markdown, while larger tables will be summarized with a link to view the full results. Code and logs will use standard Slack code blocks.
Chat artifacts are not visible via Slack.
Case context and chat continuity
When you chat with an agent while you have a case open, the agent automatically receives the case context. This allows for immediate, context-aware analysis without requiring you to manually provide case details. The agent can visualize the entire scope of the investigation, interpreting complex relationships between entities and identifying patterns across the case data.
When you begin a chat while viewing a case, the agent automatically receives the relevant case context.
Note
Case data is only loaded when you send your first message. Opening the chat interface without sending a prompt does not provide the agent with the case context.
If you have not yet sent a message while viewing a case, and you switch cases:
If you return to a case with a previous chat history, the chat and the associated context automatically load.
If no chat history exists for the case, the agent automatically opens a new chat.
If you are in the middle of a chat and switch to a different case, the Agentic Assistant asks if you want to start a new chat for the case you are viewing. If you begin a new chat and send a prompt, the case context for the new case is provided to the agent.
User action | Context status |
|---|---|
Open chat, no message sent | No context loaded. |
Send first message | Context for the current case is loaded. |
Switch cases (no active chat) | No context is transferred. Agent remains 'blank.' |
Switch cases (active chat) | Agentic Assistant suggests you start a new chat to switch the context to the new case or automatically resumes an existing chat. |
Switch to a case with chat history | Previous chat and context are automatically resumed. |
Chat navigation and system behavior
Navigate long responses: If an agent's response is long, you can jump directly to the last line of the response by clicking the anchor icon.
Start over: Sometimes an investigation takes a new direction, or you want to pivot to a different task. You can always open a new conversation or start a new investigation path with a new agent whenever needed.
Processing time: While an agent is processing a prompt, you can begin typing a new prompt. However, you can only submit this new prompt once the previous one has completed its processing. For complex actions, the system may indicate that it's taking some time. Actions exceeding five minutes result in an error.
Review the plan and execution
Cortex Agentic Assistant operates with transparency. The agent's proposed plan or steps for any action are always visible.
Click Plan and expand the chevron to review the detailed breakdown of what the agent intends to do.
JSON artifacts are created when agents create objects or retrieve information. JSON artifacts are available directly in the agent’s plan view to provide technical context for results.
Note
An agent's proposed plans and results may contain inaccuracies or errors. Always review the results carefully to ensure you fully understand the proposed action before proceeding.
Safeguards for chat security and control
Cortex Agentic Assistant implements the following safeguards to ensure agent plans and executions are secure, approved, and maintains your control over critical system changes.
Agents are designed to intelligently validate their proposed plans, ensuring that all necessary permissions are in place before any action is taken.
Cortex Agentic Assistant clarifies ambiguous prompt intentions and blocks requests that may be exploitative or harmful, for example, to perform a malicious operation.
For any sensitive actions, agents will always require your explicit approval.
Your conversations within the Agentic Assistant chat are private. However, for transparency and auditing purposes, Cortex XSIAM audit logs record all actions performed by the agents in response to your prompts. This ensures transparency by providing a detailed, traceable record of who initiated an action, what action was taken, and when, without logging the private content of your prompts themselves.
Tip
You can quickly jump to different product pages within Cortex XSIAM by typing / in the prompt area. This shortcut is a built-in navigation feature that is available even if the Cortex Agentic Assistant is disabled.