Agents
Configure Agent
Set your agent's identity, instructions, persona, knowledge, and tools across the six agent tabs.
Every agent has six tabs: Agent, Brain, Call, Actions, Procedures, and Deploy. Use this page as the reference when you set up a new agent or change how an existing one behaves.
Agents are workspace-scoped. Each agent belongs to one workspace. If you have several workspaces, create separate agents in each. See Workspaces.
Agent tabs
A live test chat panel sits on the right of the agent page, so you can test changes as you make them.
Agent tab
The Agent tab shapes how your agent talks and behaves in every conversation.
Identity
Agent name: Pick a name that reflects the agent's job. It appears in My Agents and helps you tell agents apart.
Status: Controls whether customers can reach the agent.
Instructions
Instructions are the foundation of your agent's behavior. They define:
- The agent's role and responsibilities
- Tone and communication style
- When to use tools and actions
- When to escalate to a person
Instructions have a 10,000 character limit. Write in your own words, then use Magic Prompt to turn your description into structured instructions. See Prompting Tips for guidance.
Welcome message
The first message customers see when they start a conversation. Keep it friendly, clear, and in your brand voice.
Suggested questions: Add up to four starter questions customers can tap instead of typing. They appear as bubbles in the agent chat and on the embed widget. The floating input widget shows up to three. Keep each question under 120 characters and write them the way a customer would ask.
Error message
What your agent says when it hits an error or can't help. Give customers a next step, such as how to reach your team.
Language
You can also set a voice language for voice interactions.
Persona
Persona controls how the agent sounds, on top of your Instructions.
Chat abilities
Agent Vision
Turn on Agent Vision so your agent can process images that customers send.
Use Agent Vision for:
- Refund evidence, such as photos of damaged products
- Product identification and verification
- Visual context for more accurate answers
- Document and receipt analysis
Product previews
You can also turn on generated product-in-space previews for chat and Meta. The agent uses your catalog, or uploaded reference images, to show a product in a scene.
- Each generated preview uses 3 to 15 messages. Most standard previews use 4 to 5. A single preview never costs more than 15 messages.
- Choose a catalog mode: none, an allowlist of products, or all products that have images.
- Upload up to 20 manual reference images per agent.
Confirm the usage cost when you turn on previews. Use them when customers need to see a product in context, not for ordinary image understanding.
Knowledge source
Link one Knowledge Base source to the agent so it can search your content.
The agent searches the assigned source with semantic (vector) search. The same source can be linked from more than one agent. For live information that isn't in your knowledge base, turn on Web Search.
Brain tab
The Brain tab holds the model, Temperature (from Focused to Creative), the Knowledge source, Suggested questions shown to users as quick-start prompts, and Live data connectors that sync external data into a dedicated knowledge source.
Pick the agent model under Agent mode. Link AI agent models combine several LLMs and route each request to the best one for the task.
See Models for the selection guide, temperature, and token limits.
Call tab
The Call tab holds voice settings for web voice and phone calls.
Voice Mode must be on in the Call tab before any voice interaction reaches the agent. See Phone Numbers & Calls.
Actions tab
Turn on the Actions your agent can use, and configure each one:
Procedures tab
Add Procedures when a job must follow fixed steps, such as qualifying a lead and creating a board item. Each agent also comes with system procedures it follows without you writing one.
Deploy tab
Turn on the Channels where the agent is live.
Slack isn't turned on from Deploy. Connect Slack from Integrations and assign Ally or this agent as the responder for your workspace. Only one agent answers Slack at a time. See Slack.
Test your agent
Use the live test chat panel on the right of the agent page.
Test conversations count toward your usage limits. Keep the agent on Draft while you test so your live performance metrics stay clean.
Best practices
- Start simple. Begin with a basic setup and add complexity gradually. A simple setup is easier to refine.
- Document your settings. Note why you made each choice. It helps when you troubleshoot later.
- Iterate on data. Review analytics and conversations regularly, and refine the setup.
- Test before going live. Test every change on Draft before you set the agent to Live.
- Monitor performance. Track answer quality, action usage, and customer satisfaction.
Troubleshooting
The agent isn't answering correctly
- Review the Instructions for clarity and specificity.
- Check that a knowledge source is assigned and has the right content.
- Confirm the agent model fits your use case.
- Test with simpler questions to isolate the issue.
Actions aren't working
- Confirm the actions are turned on in the Actions tab.
- Check each action's configuration for errors.
- Make sure the Instructions say when to use each action.
- Test each action in the live test chat.
Answer quality is poor
- Make the Instructions more specific.
- Add relevant content to the knowledge source.
- Adjust temperature and token limits. See Models.
- Try a different agent model, such as Link Core Smart for higher quality.