Overview
This is where knowledge agents become truly powerful. While knowledge lets your agent understand and answer questions, tools let it actually do things. Knowledge agents can orchestrate three types of tools:| Tool Type | What It Does | Best For |
|---|---|---|
| Workflow Agents | Call your existing Agent.AI workflow agents | Custom automations you’ve built |
| MCP Servers | Connect custom tools via Model Context Protocol | Advanced/developer capabilities |
| Composio Integrations | Take actions in 100+ external apps | Slack, Gmail, HubSpot, etc. |
How Function Calling Works
When you enable tools for your knowledge agent, here’s what happens:- System instructions that explain when to use each tool
- Clear tool names and descriptions
- Good conversation design
Workflow Agents as Tools
This is the most common and powerful integration. Your knowledge agent can call any of your workflow agents to accomplish tasks.How It Works
Workflow agents are your deterministic automations (step-by-step processes). Knowledge agents are conversational and decide when to invoke those automations. Think of it like:- Workflow agents = Specialized workers (do one thing really well)
- Knowledge agents = Manager (decides who to call for what task)
Adding Workflow Agents
- Navigate to your knowledge agent builder
- Click the “Action Agents” tab
- You’ll see a list of all your workflow agents
- Check the box next to each workflow you want to enable
- Click “Save Action Agents Selection”
Start with 2-3 workflows maximum when testing. Add more once you’ve verified each one works individually.
Making Your Knowledge Agent Use Workflows
Just enabling a workflow doesn’t mean your knowledge agent will use it. You need to guide the AI through system instructions.Good System Instructions for Workflow Tools
Bad example: “You have workflows available to help users.”This is too vague. The AI won’t know when to use what.
Naming Your Workflow Agents
Make sure your workflow agents have clear, descriptive names: Good names:- “Company Research Tool”
- “Email Sender”
- “Data Analyzer”
- “LinkedIn Profile Enricher”
- “Agent 1”
- “My Workflow”
- “Test”
- “Automation”
Passing Data Between Expert and Workflow Agents
When your knowledge agent calls a workflow, it can: Send data to the workflow:- User input from the conversation
- Data retrieved from knowledge base
- Results from previous tool calls
- Workflow outputs
- Structured data
- Generated content
- Status/error messages
Example: Research Assistant with Multiple Workflows
Enabled workflows:- “Web Search” - Searches the internet
- “Company Enrichment” - Gets company details
- “Create Google Doc” - Saves to Google Docs
MCP (Model Context Protocol) Servers
MCP is an advanced feature for developers who want to build custom tools for their knowledge agents.What is MCP?
Model Context Protocol is an open standard for connecting AI agents to external tools and data sources. It allows you to:- Build custom tools in any programming language
- Connect to proprietary systems
- Extend agent capabilities beyond built-in features
- Share tools across different agents
When to Use MCP
Use MCP servers when:- You need custom functionality not available in workflow agents or integrations
- You’re connecting to proprietary internal systems
- You want fine-grained control over tool behavior
- You’re comfortable with development/technical setup
Setting Up MCP Servers
Technical knowledge required: Setting up MCP servers requires development experience. Most users should start with workflow agents and Composio integrations.
- Build your MCP server following the MCP specification
- Deploy it somewhere your knowledge agent can access
- Register it with your knowledge agent:
- Go to knowledge agent settings
- Navigate to MCP configuration
- Add your server URL and authentication
- Reference it in system instructions so the agent knows when to use it
MCP vs Workflow Agents
| Aspect | Workflow Agents | MCP Servers |
|---|---|---|
| Setup difficulty | Easy (no-code) | Advanced (coding required) |
| Best for | Business automations | Custom technical tools |
| Visibility | Visual workflow builder | Code-based |
| Sharing | Clone in platform | Deploy infrastructure |
Composio Integrations
Composio provides 100+ pre-built integrations with popular apps and services. This lets your knowledge agent take actions in external systems.Available Integrations
Communication:- Slack - Send messages, create channels
- Gmail - Send emails, read inbox
- Discord - Post messages, manage servers
- HubSpot - Create contacts, deals, notes
- Salesforce - Update records, search data
- Pipedrive - Manage pipeline
- Google Drive - Create/read documents
- Notion - Update databases, create pages
- Asana - Create/update tasks
Setting Up Composio Integrations
- Navigate to your knowledge agent builder
- Click the “Integrations” tab (or “Composio” tab)
- Browse available integrations
- Click on an integration you want to enable
- Authenticate - Connect your account (OAuth flow)
- Configure permissions - Choose what the agent can access
- Save the integration
Guiding Your Agent to Use Integrations
Like with workflow agents, you need to tell your knowledge agent when to use integrations via system instructions.Example: Slack Integration
System instructions:Security Best Practices
Important: Integrations have real access to your accounts. Be careful with public knowledge agents.
- Don’t share agents that have sensitive integrations
- Review what permissions each integration needs
- Add confirmation steps for destructive actions
- Consider separate accounts for agent integrations
Combining Multiple Tools
The real power comes from knowledge agents orchestrating multiple tools together.Example: End-to-End Sales Research Agent
Enabled tools:- “Company Research” workflow agent
- “LinkedIn Profile Enrichment” workflow agent
- HubSpot integration (Composio)
- Gmail integration (Composio)
Multi-Tool Orchestration Best Practices
Do:- Let the agent decide the order of tool calls
- Build confirmation steps for sensitive actions
- Chain related tools logically (research → enrich → save)
- Use results from one tool to inform the next
- Try to hard-code exact sequences (let AI adapt)
- Enable too many tools at once (start with 3-5)
- Skip confirmation on actions like sending emails
- Forget to handle errors gracefully
Troubleshooting Tools
Knowledge agent isn't calling my workflow
Knowledge agent isn't calling my workflow
Symptoms: Agent responds conversationally but doesn’t invoke the workflowPossible causes:
- Workflow not enabled in Action Agents tab
- System instructions don’t mention when to use it
- Workflow name is unclear
- Agent doesn’t think it’s relevant to the request
- Verify workflow is checked in Action Agents tab
- Add explicit instructions: “Use [workflow name] when users ask to [task]”
- Rename workflow to be more descriptive
- Ask more directly: “Use the [workflow name] to research…”
- Test workflow independently to ensure it works
Workflow keeps failing or returning errors
Workflow keeps failing or returning errors
Symptoms: Agent calls the workflow but gets errorsPossible causes:
- Workflow itself has a bug
- Knowledge agent passing wrong data format
- Workflow expecting different inputs
- Test the workflow agent independently (run it directly)
- Check workflow input requirements
- Review what data the knowledge agent is passing
- Update system instructions to format data correctly
- Add error handling to the workflow
Agent calls the wrong tool
Agent calls the wrong tool
Symptoms: Agent uses Tool A when Tool B would be betterPossible causes:
- Tool names/descriptions are ambiguous
- System instructions unclear about when to use what
- User request was vague
- Make tool names more specific and distinct
- Add clear boundaries in system instructions: “Use Tool A for [specific case]. Use Tool B for [different case].”
- Test with specific requests that clearly need one tool
- Reduce number of similar tools enabled
Composio integration authentication failed
Composio integration authentication failed
Symptoms: Can’t connect or authenticate with external servicePossible causes:
- OAuth flow expired or interrupted
- Wrong permissions requested
- Service credentials changed
- Rate limits exceeded
- Re-authenticate the integration (disconnect and reconnect)
- Check service status (is the external service down?)
- Review required permissions for the integration
- Wait if rate limited, then try again
- Contact support if integration consistently fails
Agent calls too many tools for simple requests
Agent calls too many tools for simple requests
Symptoms: Agent over-engineers simple tasks by calling multiple toolsPossible causes:
- System instructions encourage thoroughness without boundaries
- Agent trying to be helpful but overdoing it
- Add efficiency guidelines to system instructions: “Use the minimum number of tools needed to complete the task”
- Specify when NOT to use certain tools
- Test with simple requests and iterate on prompts
- Consider if you enabled too many overlapping tools
How do I know which tool was called?
How do I know which tool was called?
Answer: As the builder testing your agent:
- Look for “[Calling workflow name…]” messages
- Watch for integration loading states
- Check agent response for explicit mentions
- Configure system instructions to announce tool usage
- Have agent explain what it’s doing
- Keep tool calls invisible for seamless experience
Advanced: Tool Call Patterns
The Research-Execute Pattern
The Pipeline Pattern
The Approval-Gate Pattern
The Fallback Pattern
Testing Your Tools
After enabling tools, thoroughly test:1. Individual Tool Testing
Test each tool separately:- “Use [workflow name] to research Microsoft”
- “Send a test message to #test-channel on Slack”
- “Create a test contact in HubSpot”
- Tool is called correctly
- Data is passed properly
- Results come back as expected
- Errors are handled gracefully
2. Multi-Tool Sequences
Test tool combinations:- “Research Company X and add them to HubSpot”
- “Analyze this data and save results to Google Sheets”
- “Find recent news and post summary to Slack”
- Tools are called in logical order
- Data flows between tools correctly
- User gets progress updates
- Final result is complete
3. Edge Cases
Test failure scenarios:- What happens if a workflow fails?
- What if an integration is disconnected?
- What if the user provides incomplete information?
- Graceful error messages
- Agent asks clarifying questions
- Doesn’t get stuck in loops
- Offers alternatives
4. Approval Workflows
Test confirmation flows:- Does agent ask before sensitive actions?
- Can user say “no” and agent respects it?
- Does agent re-confirm if request changes?
Best Practices Summary
Tool Integration Best Practices
DO:
- Start with 2-3 tools and add more gradually
- Write explicit system instructions for each tool
- Use clear, descriptive tool names
- Test each tool individually before combining
- Add confirmation steps for sensitive actions
- Let the AI decide when to use tools (don’t hard-code)
- Enable every tool at once
- Assume the AI knows when to use tools without guidance
- Skip testing multi-tool scenarios
- Give public agents access to sensitive integrations
- Forget to handle errors and edge cases
Next Steps
Now that you understand how to give your knowledge agent powerful action-taking capabilities:Manage Conversations
Learn about conversation management, sharing, and user experience
Best Practices
Advanced techniques for building exceptional knowledge agents
Troubleshooting
Solve common issues and optimize performance
Build Workflow Agents
Create the workflow agents your knowledge agent will call
Remember: Tools transform your knowledge agent from a conversational assistant into a powerful automation orchestrator. Start simple, test thoroughly, and gradually build up to complex multi-tool workflows.

