Agent Configuration Guide
1. Overview
Agents are the core execution units in FiraClaw. Each agent can be assigned different "brains" (large models), different "toolboxes" (MCP and toolsets), and different "professional capabilities" (Skills).
By configuring agents with different expertise, you can build an AI team:
- CEO Agent: Team manager, responsible for scheduling and decision-making.
- Python Expert: Specializes in coding and data processing.
- Copywriting Assistant: Specializes in polishing articles and writing emails.
- Data Analyst: Specializes in reading Excel and generating charts.
This manual will guide you on how to create, configure, and manage these agents.
2. Interface Layout Introduction
Access path: Click the Agent icon in the left menu.
Area Names and Functions
| Number | Area Name | Function Description |
|---|---|---|
| ① | Agent List | Displays all agents you have created. CEO Agent has a special marker. |
| ② | Marketplace Entry | Switch to Agent Marketplace to download officially prepared professional agents. |
| ③ | New Button | Manually create a brand new agent. |
| ④ | Search and Filter | Quickly find agents by specific name or tag. |
| ⑤ | Operation Menu | Perform operations such as editing, copying, and deleting on the selected agent. |
3. Manually Creating Agents
3.1 Basic Information Settings
Click the New Agent button to enter the configuration page.
| Configuration | Description | Example |
|---|---|---|
| Name | Easy-to-identify name. CEO will match based on name when scheduling. | Python Code Expert |
| Description | Detailed description of the agent's expertise. This description is very important as CEO will read it to decide whether to hire this agent. | Good at writing Python scripts, proficient in web crawlers, data analysis, and file automation processing. |
| Avatar | Customizable icon for quick identification. | Upload a Python Logo. |
| Welcome Message | The first guide message displayed when switching to this agent in AI Assistant dialogue. | Hello, I am your Python assistant. Please tell me what code functionality you need to write? |
3.2 Model Configuration (Required)
Each agent needs to be assigned a "brain" to think. You can assign models with different capabilities to different agents.
- Select Existing Model: Choose a model already created on the Model Configuration page from the dropdown list.
- Configuration Strategy:
- High logic requirements (coding) → Recommended to configure Claude or GPT-4 type models.
- Simple repetitive tasks (format conversion) → Recommended to configure DeepSeek or GPT-3.5 type models to save costs.
Kind Reminder: If the dropdown list is empty, please go to Model Configuration in the left menu first to add at least one default model.
3.3 Capabilities Extension Configuration (Optional, but Recommended)
This is the key step to make agents "capable".
| Configuration | Function | Operation Method |
|---|---|---|
| MCP Services | Endow agents with the ability to connect external data or operate software. | Check MCP already created or downloaded from marketplace (such as Oculos desktop operation). |
| Skills | Endow agents with specific workflow knowledge or SOPs. | Check Skills already created or downloaded from marketplace (such as PDF processing). |
| Toolset | Control which system-level tools the agent can use (such as reading/writing files, sending network requests). | Select a preset toolset. If left blank, all system tools are open by default (higher permissions, use with caution). |
3.4 Advanced Settings
- Max Reasoning Steps: Limit how many steps an agent can think when solving a problem to prevent infinite loops. It is recommended to keep the default.
- Temperature Parameter: Control output randomness. For code generation, it is recommended to lower it (0.1-0.3); for creative writing, it can be higher (0.7-0.9).
4. CEO Agent Special Instructions
The system-built CEO Agent has a special status, and its configuration items are slightly different from other agents:
- Cannot be Deleted: CEO is the system's core scheduler.
- Model Configuration: CEO needs strong logical planning ability. It is recommended to configure the currently strongest model.
- Description Cannot be Modified: CEO's description is the system-preset scheduling logic.
- Toolset Restrictions: CEO has system-level permissions by default such as scheduling other agents and creating tasks. These permissions are not open to ordinary agents.
5. Agent Marketplace
5.1 Browsing and Downloading
- Click the Marketplace tab at the top of the agent page.
- Browse officially provided premium agents (such as "Xiaohongshu Viral Writer", "English Paper Polisher").
- Click the Download button, and the agent will be automatically added to your list.
5.2 About Uploads and Customization
- Security Statement: To ensure the software environment security for all users, the Agent Marketplace does not open public uploads temporarily.
- Customization Needs: If you need to batch create agents for specific industries or want official to develop certain types of agents, please submit your needs through the in-software Feedback & Suggestions function. We will contact you.
6. Editing and Deleting Agents
- Edit: In the agent list, hover the mouse over the agent card and click the Edit icon. After modification and saving, the new configuration will take effect in the next dialogue.
- Copy: If you want to fine-tune and create a new version based on an existing agent, you can use the Copy function to quickly clone.
- Delete: Deleting an agent will not delete the task records that the agent has executed. Note: CEO Agent cannot be deleted.
7. Best Practices Suggestions
| Scenario | Recommended Configuration |
|---|---|
| Daily chat / Coordination | Use CEO Agent directly with the strongest model configured. |
| Coding / Data Processing | Create CodeMaster agent, configure a large model with strong coding ability, enable file read/write toolset. |
| Operating Office Software | Create Office Assistant agent, bind Oculos MCP desktop operation service. |
| Translation / Polishing | Create Linguist agent, configure long-context model, load translation polishing Skill. |
8. FAQ
Q: Why did I create a "coding" agent, but when I switch to it in the dialogue, it says it can't code?
A: Please check the Description field of that agent. The large model will first read the description to confirm its identity. Please ensure the description contains clear instructions like "You are a Python programming expert".
Q: Can I let CEO Agent use a cheaper model?
A: Yes, but it is strongly not recommended. CEO needs to handle complex logical decomposition and task distribution. A weak model may lead to scheduling failures or task understanding errors. It is recommended to use the strongest model for CEO, and execution labor "employee" agents can use cost-effective models.
Q: I configured MCP for an agent, but it doesn't use that software during dialogue. What should I do?
A: Please explicitly prompt it has that tool in the agent's description. For example: "You can use Oculos MCP to control mouse and keyboard operations on Windows desktop."
Q: If I modify an agent's configuration, will it affect tasks that are currently running?
A: No. Tasks that have been triggered and are currently executing will continue using the agent configuration snapshot from when the task started. New configurations only apply to dialogues and tasks created after the configuration modification.