Your First Agent
A deeper walkthrough of creating a single agent from scratch. You will learn how to craft an effective system prompt, choose the right model, assign tools, and test the result -- all from the Dashboard.
Before You Start
Make sure you have:
- A workspace and team created (see the Quickstart)
- At least one LLM key saved in Settings > LLM Keys
Step 1: Plan the Agent
Before opening the Dashboard, think through what this agent should do:
| Question | Your decision |
|---|---|
| What is its purpose? | Example: Answer IT support questions |
| What personality should it have? | Example: Helpful, patient, technically accurate |
| What should it NOT do? | Example: Never access production systems directly |
| When should it escalate? | Example: When the issue requires physical access or admin credentials |
Writing this down first makes the system prompt much easier to write.
Step 2: Create the Agent
- Navigate to your team page
- Click Add Agent
- Fill in the identity fields:
| Field | What to enter | Tips |
|---|---|---|
| Name | A clear, recognizable name | Use a role-based name like "IT Helpdesk Assistant" rather than a generic one |
| Role | The agent's job title | This appears in team views and handoff logs |
| Description | A sentence about what the agent does | Helps colleagues find the right agent quickly |
Step 3: Write the System Prompt
The system prompt is the most important part of your agent. It defines personality, knowledge, capabilities, and boundaries. A well-structured prompt has four sections:
Section 1 -- Identity and Role
Tell the agent who it is and what it does in plain language.
You are the IT Helpdesk Assistant for Acme Corporation. You help
employees resolve technical issues and answer IT-related questions.
Section 2 -- Capabilities and Boundaries
Be explicit about what the agent can and cannot do. This prevents the agent from overstepping.
Your capabilities:
- Answer questions about company software and systems
- Guide users through troubleshooting steps
- Create support tickets for complex issues
- Look up IT policies and procedures
Your boundaries:
- You cannot directly access production systems
- You cannot make changes to user accounts without approval
- You cannot access confidential employee data
- For security-sensitive requests, always verify identity first
Section 3 -- Domain Knowledge
Give the agent the information it needs to do its job well.
Company IT environment:
- Email: Google Workspace (Gmail, Calendar, Drive)
- Communication: Slack
- VPN: Cisco AnyConnect
- SSO: Okta (all apps use single sign-on)
- Ticketing: ServiceNow
Section 4 -- Communication Style
Define how the agent should talk to people.
Communication guidelines:
- Be friendly but professional
- Use clear, step-by-step instructions
- Avoid jargon unless the user uses it first
- Confirm understanding before proceeding
- Always ask if the issue is resolved before closing
A system prompt does not need to be long. Aim for clarity over length. An agent with a concise, well-structured prompt will outperform one with pages of instructions.
Paste your complete prompt into the System Prompt field in the agent editor.
Step 4: Choose a Model
Click the Model dropdown to select which LLM powers this agent. Your choice depends on the task:
| Model | Best for | Speed | Notes |
|---|---|---|---|
| Claude Sonnet (latest) | General tasks, balanced performance | Fast | Great default choice for most agents |
| Claude Opus (latest) | Complex reasoning, nuanced responses | Slower | Best for agents that handle ambiguous or high-stakes tasks |
| Claude Haiku (latest) | Simple tasks, high volume | Fastest | Good for triage agents that route rather than solve |
| GPT-4.1 | General tasks (OpenAI ecosystem) | Fast | Requires an OpenAI key |
| Gemini 2.5 Pro | Long-context analysis | Fast | Requires a Google key |
Since MeetLoyd is Bring Your Own Key, you can mix models across agents. Use a fast model for triage and a powerful model for complex resolution -- all within the same team.
Step 5: Assign Tools
Tools give your agent capabilities beyond conversation. From the agent editor, click the Tools tab.
| Tool category | Examples | When to use |
|---|---|---|
| Memory | Remember user preferences, past issues | When the agent needs context across conversations |
| Integrations | Google Workspace, Microsoft 365, Slack | When the agent needs to read or write data in other systems |
| MCP Servers | Custom tools via Model Context Protocol | When you have internal tools the agent should use |
| Skills | Pre-built expertise packages from the Store | When you want domain-specific behavior without writing prompts |
To add a tool:
- Click Add Tool
- Browse or search for the tool you need
- Configure any required settings (scopes, allowed domains, etc.)
- Click Save
Agents perform better with a focused set of tools. Add only what is needed for the agent's defined role. You can always add more later.
Step 6: Test Your Agent
- Save the agent and go back to your team page
- If the team is not yet running, start it using the Team Starting Wizard
- Switch to Openspace and open a conversation with your agent
- Test with realistic scenarios:
Simple question -- Does the agent answer correctly based on its domain knowledge?
Boundary test -- Ask it to do something outside its boundaries. Does it refuse appropriately?
Escalation test -- Present a complex issue. Does it offer to escalate or create a ticket?
Memory test (if Memory tool is assigned) -- Tell it something, then in a later message, reference what you said. Does it remember?
Step 7: Refine
Based on your testing, go back and adjust:
| What to change | Where to change it |
|---|---|
| Agent is too verbose or too terse | Edit the communication style section of the system prompt |
| Agent oversteps its boundaries | Make the boundaries section more explicit |
| Agent lacks domain knowledge | Add more context to the domain knowledge section |
| Agent is too slow | Switch to a faster model (Haiku or Sonnet) |
| Agent cannot perform needed actions | Add the appropriate tools |
You can edit the agent at any time from the team page without restarting the team.
What You Have Built
You now have a properly configured agent with:
- A clear identity and role
- Defined capabilities and boundaries
- Domain knowledge relevant to its job
- The right model for the task
- Tools to interact with external systems