How to create an Agent

Before you start

Agents are a Power-tier feature. If you don't see "Agents" in your sidebar, you're on Free, Pro, or Pro+ — upgrade to Power to unlock them.

What an Agent is

An Agent is a reusable AI assistant with three things baked in:

• A system prompt — the instructions that shape how it responds (e.g., "You are a contract reviewer. Flag risky clauses in plain English.").
• Tools — what it can DO (web search, code execution, file access, MCP integrations).
• Optional knowledge files — PDFs or docs the agent can reference.

Once saved, you can use the agent like a regular chat — but it always behaves the way you configured it.

Step-by-step

1. Click Agents in the left sidebar (the grid icon near the top).
2. Click Create Agent.
3. Give it a clear name — e.g., "Email drafter," "SOC 2 reviewer," "Daily standup helper." This is what you'll see in the dropdown when picking which agent to use.
4. Write the system prompt. Be specific. "Write professional emails" is too vague. "You write friendly, brief emails for B2B sales follow-ups. Always end with a single clear question." is better.
5. Pick the model. Power tier has access to all four — pick the cheapest that does the job. Lite for simple tasks, Plus for long context (1M tokens), Pro for the hardest reasoning.
6. (Optional) Enable tools: web search if the agent needs current info, code execution if it needs to run scripts.
7. (Optional) Upload knowledge files — the agent can search them when answering.
8. Click Save. Done.

Using your agent

From the chat input, click the model picker and choose your agent. Type a message — it'll respond using the system prompt, tools, and knowledge you set up. You can switch back to a regular model at any time.

Examples that work well

• Daily standup helper — system prompt asks for yesterday/today/blockers, formats the output, files it.
• Code reviewer — paste a diff, the agent reviews against your standards (defined in the system prompt).
• Writing assistant with your style — upload three samples of your writing as knowledge files; the agent matches your voice.
• Customer-support draft writer — agent reads your help center docs (knowledge files) and drafts responses to incoming tickets.

Common mistakes

• System prompt too short — vague prompts produce vague responses. Spend time on this.
• Tools you don't need — turning on every tool slows the agent down and costs more tokens. Enable only what's necessary.
• Wrong model — Pro for everything is wasteful. Lite is fine for templated tasks.