Skip to main content
Building a custom agent takes a few minutes. You give it a name, write a system prompt that defines its behavior, choose a model, and configure which tools and apps it can use. Once saved, you can chat with it anytime from /agents.

​
Steps to Create an Agent

1

Navigate to /agents

Click Agents in the topbar, or go directly to /agents.
2

Click New Agent

Click the New Agent button or the + icon. The agent creation form opens.
3

Enter a name and description

  • Name: What you want to call the agent. This appears in the agent list and in chat. Example: “TypeScript Dev”, “Research Analyst”, “Writing Coach”.
  • Description: A short summary of what this agent does. Helps you remember its purpose when you have multiple agents.
4

Choose an avatar or icon (optional)

Select an icon or upload an image to visually distinguish this agent from others on your /agents page.
5

Select a base model

Choose the AI model that will power this agent. Different models suit different use cases:
  • Fast models — best for quick Q&A, drafting, routine tasks
  • Reasoning models — best for complex analysis, debugging, multi-step logic
  • Balanced models — good general-purpose default for most agent types
6

Write the system prompt

The system prompt is the most important part of your agent. It defines the agent’s role, knowledge, tone, and behavior. See examples below.
7

Enable tools

Toggle on the tools this agent should have access to: web search, deep research, canvas, file uploads, voice, and others.
8

Select connected apps

Choose which of your connected OAuth apps this agent can use. Only apps you’ve already connected in Settings → Connectors are available.
9

Save

Click Save or Create. The agent appears on your /agents page. Click it to start a chat.

​
System Prompt Examples

​
TypeScript coding assistant

You are a senior TypeScript developer. Help me write clean, typed, well-documented code.

Guidelines:
- Always include proper TypeScript types — avoid implicit any
- Use functional patterns over class-based when possible
- Prefer named exports
- Suggest best practices proactively
- When reviewing code: point out bugs first, then style improvements
- Use concise, technical language — I'm an experienced developer

Stack context: React 18, TypeScript 5, Tailwind CSS, Supabase, Zod for validation.

​
Market research analyst

You are a market research analyst specializing in SaaS products and competitive intelligence.

When researching:
- Always cite sources with URLs when available
- Structure responses with: Summary, Key Findings, Data Points, Recommendations
- Focus on actionable insights, not just facts
- Flag when data is uncertain or outdated
- Target audience: product managers and founders making strategic decisions

​
Writing coach

You are an experienced writing coach and editor. Your goal is to make my writing clearer,
more compelling, and more concise.

Approach:
- Lead with the most important feedback
- Give specific rewrites, not just general suggestions
- Explain why a change improves the writing
- Match my voice — don't replace it with yours
- For long pieces: start with structure and argument before line edits

​
Tips for a Great System Prompt

Define the role clearly. A specific role (“You are a senior TypeScript developer specializing in React”) is better than a vague one (“You help with coding”). Specify the output format. If you always want bullet points, code blocks, or specific sections, say so in the prompt. Add relevant background context. Include your stack, your style preferences, your audience — anything that helps the agent give consistently relevant responses. List constraints explicitly. “Only answer questions about this codebase — redirect unrelated questions” keeps the agent focused on its purpose. Keep it under 1,000 words. Very long system prompts consume context window space. Write comprehensive but efficient instructions.

​
After Creating

Click the agent on /agents to start a chat. Test it with a few representative tasks. If the output isn’t quite right, go back and edit the system prompt — most agents need a few iterations to reach ideal behavior. See Configure an Agent for editing guidance.