Skip to main content

​
How Conversations Work

When someone interacts with your knowledge agent, each interaction creates a conversation. Think of conversations like chat threads - they have:
  • History - All messages in the conversation
  • Context - The AI remembers previous messages
  • Auto-generated titles - AI creates descriptive names
  • Shareable links - Can be shared publicly
  • Forking - Others can copy and continue from any point
Each conversation is isolated - what happens in one doesn’t affect others.

​
Conversation Lifecycle

User opens your knowledge agent
        ↓
New conversation created automatically
        ↓
User and agent exchange messages
        ↓
AI generates a descriptive title
        ↓
Conversation saved to history
        ↓
User can share, fork, or continue later

​
Auto-Generated Titles

Your knowledge agent automatically generates descriptive titles for conversations after the first few messages. How it works:
  • AI analyzes the conversation topic
  • Generates a concise, descriptive title
  • Title appears in conversation history
  • Makes it easy to find past conversations
Examples of auto-generated titles:
  • “Researching AI automation competitors”
  • “Creating social media campaign for product launch”
  • “Analyzing Q3 sales data and trends”
  • “Debugging authentication module errors”
You can’t customize the title format, but you can influence it through:
  • Clear user requests (AI titles based on main topic)
  • Focused conversations (don’t jump between unrelated topics)
If a conversation covers multiple topics, the AI typically titles it based on the first major topic discussed.

​
Conversation History

All conversations are automatically saved and accessible from the conversation history panel.

​
Accessing Your Conversations

As the agent builder:
  1. Open your knowledge agent
  2. Look for the conversation list (usually left sidebar)
  3. See all conversations sorted by most recent
  4. Click any conversation to reopen it
For users interacting with your agent:
  1. Conversations are saved in their account
  2. They can return and continue conversations
  3. History persists across sessions

​
What’s Saved

Each conversation stores:
  • All messages (user and agent)
  • Tool calls made
  • Knowledge retrieved
  • Timestamps
  • Auto-generated title
Privacy note: Builders can see conversations with their own knowledge agents. Consider this when deciding what features to enable on public agents.

​
Sharing Conversations

One of the most powerful features of knowledge agents is the ability to share individual conversations via public links.

​
How to Share a Conversation

  1. Have a conversation with your knowledge agent
  2. Look for the share icon (usually top-right of conversation)
  3. Click to generate a shareable link
  4. Copy and share the link anywhere
The link looks like: https://agent.ai/chat/[conversation-id] When someone opens a shared conversation link, they see:
  • Full conversation history - All messages in the conversation
  • Read-only view - They can read but not modify the original
  • Fork option - They can copy the conversation and continue it
  • Agent information - Who built it, description

​
Use Cases for Sharing

Showcase examples: 
Share great examples of your knowledge agent in action:
- Marketing campaigns it created
- Research it conducted
- Code it helped debug
- Reports it generated

Use these as portfolio pieces or demos.
Collaborative work: 
Work with someone on a project:
- Start conversation with your knowledge agent
- Get to a point where you want input
- Share link with colleague
- They can fork and continue
Support and troubleshooting: 
If something isn't working:
- Create a conversation showing the issue
- Share with support or the agent builder
- They can see exactly what happened
Teaching and examples: 
Create example conversations showing:
- How to use the agent effectively
- What kinds of questions to ask
- Sample workflows end-to-end

Share these as tutorials.

​
Privacy Considerations

Important: Shared conversation links are public - anyone with the link can view the conversation.Don’t share conversations containing:
  • Personal information (emails, phone numbers, addresses)
  • Confidential business data
  • API keys, passwords, or credentials
  • Private customer information
  • Sensitive internal discussions
Before sharing, review the entire conversation to ensure nothing sensitive is included.

​
Best Practices for Sharing

Do:
  • Review the conversation before sharing
  • Share conversations that demonstrate value
  • Use as examples in documentation
  • Share success stories and use cases
  • Include context when sharing (explain why it’s interesting)
Don’t:
  • Share sensitive information
  • Share conversations with errors if showcasing capabilities
  • Share incomplete conversations that might confuse viewers
  • Assume shared links are private (they’re public)

​
Forking Conversations

Forking lets users copy a conversation and continue it themselves. This creates powerful collaboration and learning opportunities.

​
How Forking Works

Original conversation (shared by builder)
        ↓
User clicks "Fork" or "Continue this conversation"
        ↓
Exact copy created in user's account
        ↓
User can now continue the conversation
        ↓
Original conversation unchanged

​
When Users Might Fork

To build on examples: 
Builder shares: "Here's how to research competitors"
User forks: Continues with their own competitor list
To customize for their needs: 
Builder shares: "Campaign strategy for SaaS product"
User forks: Adapts strategy for their specific product
To learn and experiment: 
Builder shares: "Complex data analysis workflow"
User forks: Tries with their own data
To collaborate asynchronously: 
Team member 1 starts conversation
Shares link
Team member 2 forks and continues
Shares updated version back

​
Enabling Productive Forking

As a builder, you can encourage forking by: Creating “template” conversations:
  • Start a conversation with your agent
  • Walk through a complete workflow
  • Stop at a point where users can customize
  • Share with instruction: “Fork this and add your data”
Example: 
Title: "Competitor Research Template"

Conversation:
Agent: "I'll help you research competitors. What industry are you in?"
[Builder]: "SaaS"
Agent: "Great! I'll research SaaS competitors. Which companies should I analyze?"
[Builder]: "Add your companies here →"

[Share this - users fork and replace with their companies]
Building progressive examples:
  • Share multiple conversations showing progression
  • Each one builds on the previous
  • Users can fork at any stage
  • Creates learning pathways

​
Managing Conversations as a Builder

​
Testing Your Agent

As you build and refine your knowledge agent, you’ll have many test conversations: Organizing test conversations:
  • Use consistent naming in your test prompts
  • Delete obviously failed test conversations
  • Keep successful examples to share later
  • Archive old tests after major updates
Starting fresh tests:
  • Always start a new conversation for each test scenario
  • Don’t reuse old conversations (context bleeds through)
  • Test with realistic user scenarios

​
Monitoring Usage

Check your conversation history to understand:
  • What users are asking for
  • Where the agent succeeds
  • Where it gets confused
  • What workflows are most popular
Use this feedback to:
  • Refine system instructions
  • Add relevant knowledge
  • Enable additional tools
  • Update sample questions
Pro tip: Review your first 10-20 real user conversations carefully. They’ll reveal assumptions you made that users don’t share, and unexpected use cases you didn’t anticipate.

​
User Experience Best Practices

​
Setting Expectations in the First Message

Your welcome message is critical. It should: Be clear about capabilities: 
Good:
"Hi! I can research companies, enrich with LinkedIn data,
and add them to HubSpot. What would you like to research?"

Bad:
"Hello! How can I help you today?"
Show example interactions: 
Include in your welcome message:
"Try asking me:
- 'Research TechCorp and its competitors'
- 'Find 10 AI startups in San Francisco'
- 'Enrich this list with funding data'"
Set boundaries: 
"I specialize in company research and CRM enrichment.
For general questions, I recommend [alternative]."

​
Conversational Flow

Acknowledge long-running tasks: 
Bad:
[Agent calls tool, user sees nothing for 30 seconds, then results]

Good:
"Let me research that for you..."
[Calls tool]
"Found 15 companies, now enriching with LinkedIn data..."
[Calls tool]
"Analysis complete! Here are the results:"
Ask clarifying questions early: 
User: "Research competitors"

Good agent response:
"I'd be happy to research competitors. A few questions:
- What industry or product category?
- Geographic focus?
- Should I include indirect competitors too?"

Bad agent response:
"Okay, researching competitors..."
[Doesn't know what to research]
Confirm before sensitive actions: 
Good:
"I've drafted an email to the CEO. Here's what I'll send:
[Shows email]
Should I send this?"

Bad:
"Email sent to CEO."
[User had no chance to review]

​
Error Handling

When things go wrong, your agent should: Explain what happened: 
Good:
"I tried to call the Company Research workflow but it returned
an error: 'API rate limit exceeded'. This means we've made too
many requests. I can try again in a few minutes, or we can
approach this differently. What would you prefer?"

Bad:
"Error occurred."
Offer alternatives: 
"The LinkedIn enrichment tool isn't responding. I can:
1. Try a different enrichment source
2. Continue with the data we have
3. Wait and try again later

What works best for you?"
Don’t get stuck: 
If a tool fails repeatedly, don't keep trying.
Agent should: "This tool seems to be having issues. Let me try
a different approach..." or "I'll skip this step for now..."

​
Conversation Analytics

While you can’t export conversation data directly, you can learn from patterns:

​
What to Look For

Common question patterns:
  • Are users asking for things your agent can’t do?
  • → Consider adding new tools or knowledge
Where conversations succeed:
  • Which workflows work smoothly?
  • → Highlight these in examples
Where conversations fail:
  • Where does the agent get confused?
  • → Update system instructions or add knowledge
Unexpected use cases:
  • Are users doing things you didn’t anticipate?
  • → Consider optimizing for these patterns

​
Iterating Based on Conversations

Weekly review process:
  1. Review 10-20 recent conversations
  2. Identify 3 common issues
  3. Make 1-2 specific improvements
  4. Test improvements with new conversations
  5. Repeat
Example improvement cycle: 
Week 1 observation: Users often ask for data exports
Action: Enable Google Sheets integration

Week 2 observation: Agent doesn't explain what it's doing
Action: Update system instructions to narrate actions

Week 3 observation: Users confused about what agent can do
Action: Update welcome message and sample questions

​
Advanced: Conversation Handoffs

For complex agents, you might want to design conversation handoffs:

​
Handing Off to Humans

System instructions:
"If the user requests something outside your capabilities,
offer to connect them with a human:

'This requires human expertise. I can:
1. Summarize our conversation so far
2. Send a notification to [team/person]
3. Save our discussion for their review

What would you prefer?'"

​
Handing Off to Other Agents

System instructions:
"For [specific task type], suggest forking to the
specialized agent:

'For advanced data analysis, I recommend forking this
conversation to our Data Analysis Knowledge agent:
[link]. It has specialized tools for [capability].
Should I prepare a summary to start there?'"

​
Troubleshooting Conversations

Symptoms: Conversations disappear or don’t persistPossible causes:
  1. Browser cookies/storage disabled
  2. Incognito/private browsing mode
  3. Account authentication issues
Solutions:
  • Ensure user is logged in
  • Check browser allows cookies and local storage
  • Try a different browser
  • Clear cache and reload
Symptoms: Agent forgets what was discussed earlierPossible causes:
  1. Conversation is very long (approaching token limits)
  2. User jumped between multiple unrelated topics
  3. Technical issue with conversation state
Solutions:
  • Start a new conversation for new topics
  • Keep conversations focused on one main task
  • If very long conversation, fork and continue fresh
  • This is rare - if it happens often, report to support
Answer:
  1. Find the conversation in your history
  2. Look for delete/trash icon (usually hover or right-click)
  3. Confirm deletion
Note: Deleted conversations cannot be recovered. If you shared the conversation link, it will no longer work.
Answer: No, conversations are immutable. You cannot edit messages after they’re sent. If you made a mistake:
  • Continue the conversation with clarification
  • Start a new conversation
  • Fork the conversation at an earlier point
This ensures shared conversations remain truthful representations.
Answer: There’s no built-in export feature currently, but you can:
  • Copy and paste the conversation text
  • Take screenshots
  • Share the link and reference it externally
  • Use the conversation as training data (upload as knowledge)

​
Next Steps

Now that you understand conversation management and sharing:

Best Practices

Learn advanced techniques for building exceptional knowledge agents

Troubleshooting

Solve common issues and optimize your agent’s performance

Configuration

Review how to write better system instructions and prompts

Tools Integration

Give your agent more capabilities to create better conversations
Remember: Every conversation is an opportunity to learn what works and what doesn’t. Review conversations regularly, share your best examples, and continuously refine based on real usage.