AI AGENT
Send one prompt and the AI plans, executes, and verifies. The ReAct pattern lets it run code, generate files, search the web, and complete research tasks end-to-end.
The AI Agent uses the ReAct (Reasoning + Acting) pattern, alternating between thinking, acting, and observing to handle complex tasks. A cloud Linux sandbox sits behind it — Python code is generated, executed, and the next step is chosen from the result, all in one conversation. The same agent can also be invoked from real-time character chat as a background sub-agent, so the voice conversation keeps flowing while heavier research or code execution runs behind the scenes. Free tier on Gemini Flash-Lite (10/day); top-tier models (Claude / GPT / Gemini Pro, etc.) on paid plans.
Given a goal, the agent breaks it down into steps and orders them itself.
Writes Python code, runs it instantly in a cloud Linux sandbox, and folds results plus generated files back into the conversation.
Combines custom tools (web search, API calls, etc.) to complete the task.
Thought / Action / Observation streams live so you can see exactly what the agent is doing.
During real-time voice chat, the character can hand off heavy work to the agent in the background — they reply briefly, and the deeper result is appended to the chat once it's ready.
Launch "AI Chat" from the sidebar. The agent lives as a mode inside AI Chat.
Toggle the ReAct mode switch at the top of the chat screen to switch from normal chat to the agent.
Choose the AI provider (Anthropic / OpenAI / DeepSeek, etc.) and the custom tools the agent can call.
The agent plans, executes, and formats results on its own. Thought / Action / Observation streams live so you can see what it does.
The AI agent shines when combined with AI Chat, custom tools, and the Linux runtime — that's where it does real work.
Custom functions the agent can call. Web search, API integration, calculators — extend the actions available to the agent.
Learn about tools →A planned standalone browser-accessible Linux sandbox, separate from the E2B sandbox that backs the ReAct agent today. Still in the works.
Learn about runtime →The agent can also be spawned in the background from real-time voice chat via a {{AGENT: …}} marker, so the conversation never blocks while the heavy work happens.
Learn about character chat →Agent mode is a mode inside AI Chat. Switch between normal chat and agent mode seamlessly in the same screen.
Learn about AI Chat →Free tier covers Flash-Lite at 10 runs/day. Experience handling complex tasks with one prompt — upgrade to a paid plan for top-tier models.