You can create custom agents tailored to your specific needs. Define a personality, pick the right tools, set sandbox restrictions, and give the agent custom instructions.

Creating an agent

Ask the Assistant to create one for you:

• "Create an agent called 'Content Writer' that specializes in writing blog posts and social media content"
• "Set up a new agent for code review that focuses on TypeScript best practices"
• "Make a research agent that only searches academic sources"

The Assistant creates the agent, and you can start using it immediately from the sidebar.

Agent settings

Once created, you can configure every aspect of an agent through the settings page. The settings are organized into sections.

Profile

The profile section defines who the agent is.

Field	Description
Avatar	Upload an image or use the default
Enabled	Toggle the agent on or off
Name	Display name shown in the sidebar and conversations
Specialty	What the agent does. Other agents read this when deciding who to delegate to, so write it clearly.
Creature	What kind of entity it is (e.g., AI assistant, robot, familiar)
Vibe	The overall feel (e.g., sharp, warm, chaotic, calm)
Personality	Character traits (e.g., funny, blunt, poetic, sarcastic)
Style	Communication style (e.g., casual, formal, terse, elaborate)
Emoji	An emoji that represents the agent

The specialty field is especially important. A good specialty like "Handles customer support inquiries using the knowledge base and CRM" helps other agents delegate effectively. A vague one like "Support agent" doesn't.

Model

Choose which LLM powers the agent by selecting a model group. Each model group defines a provider, model, fallbacks, temperature, and token limits.

Model group	Best for
primary	General conversation and coordination
coding	Code generation and technical tasks
reasoning	Deep analysis, research, and complex reasoning

You can customize model groups in the models section of your config file.

Instructions

The prompt field contains the agent's system instructions. This is where you define the agent's behavior, workflows, and constraints. Every agent starts with a default prompt. You can customize it or replace it entirely.

Use this to:

• Define specific workflows or processes
• Set communication style or tone
• Restrict what the agent should or shouldn't do
• Provide domain-specific context

Tools

Toggle which tool groups the agent has access to. Tools are organized by category, and each group has a description of what it enables. Only include tools the agent actually needs. A research agent doesn't need CLI access. A coding agent doesn't need voice.

See What Agents Can Do for the full list.

Skills

Control which installed skills the agent can use. By default, agents have access to all installed skills. You can restrict to specific skills or disable skill access entirely.

See Extending Agents with Skills for more.

Sandbox

The sandbox section controls the agent's execution environment. Every agent runs inside a sandbox that restricts what it can access.

Resource limits:

Setting	Description
Max Memory (%)	Maximum memory the agent can use, as a percentage of system memory
Max CPU (%)	Maximum CPU the agent can use, as a percentage of available cores
Timeout (seconds)	Maximum execution time per command (0-300)

Network:

Setting	Description
Network Access	Toggle to allow or block all network requests
Allowed Destinations	Specific hostnames, IPs, or CIDR ranges the agent can reach. Supports formats like www.example.com, 1.1.1.1, 10.10.100.1/24

Shared files:

Setting	Description
Shared Paths	Files or directories the agent can read inside the sandbox. You can type paths manually or browse the file system.

See Sandbox for more on security and isolation.

Credentials

View and manage vault access grants for the agent. This section shows which secrets the agent has been granted access to, including the vault provider, expiration date, and environment variable prefix. You can revoke grants from here.

See Managing Secrets & API Keys for more.

Tips

• Start simple. Create an agent with just a name, specialty, and prompt. Add tools and sandbox settings as needed.
• Write good specialties. Other agents use the specialty to decide when to delegate. Be specific about what this agent handles.
• Use the right model group. reasoning for research and analysis, coding for development tasks, primary for everything else.
• Limit tools. Don't give every agent every tool. Fewer tools means less noise in the prompt and better token efficiency.
• Lock down the sandbox. If an agent only needs to access one API, restrict network destinations to that host.

Next steps

• Agent Types to see how the built-in agents are configured
• Delegation & Tasks for how agents work together
• Sandbox to understand security and isolation