Amp expands AI agent capabilities beyond coding using OpenAI models

Amp, a startup building frontier AI agents for developers, is leveraging OpenAI models to expand agent utility beyond code generation into design, marketing, and real-time interaction. CEO Quinn Slack notes that OpenAI models account for the majority of tokens used in Amp, citing transparency and consistency as key factors in their adoption.

Sign in with ChatGPT and Token Costs

To address token cost constraints, Amp launched a Sign in with ChatGPT integration. This feature provides users with access to more tokens and models, including GPT-6 Astra, encouraging users to assign larger tasks to agents without micromanaging usage. Slack highlights that some Amp users consume 10 billion tokens a week, and removing the need to count every token allows agents to demonstrate their full potential in planning and post-ship work.

GPT-6 Astra for Design and Creation

Amp’s internal team uses GPT-6 Astra for high-quality writing and design, including marketing materials and documentation. The model helps maintain Amp’s specific design and writing standards. Slack also demonstrated the model's versatility by using it to create a 3D model of his house in Blender, which his children then used to build games.

Real-Time Control with GPT-Live-1

Amp introduced Puck, a conversational AI assistant built into the platform that helps users launch and manage other agents. Powered by GPT-Live-1, the "Calling Puck" feature allows users to interact with Puck in real time, even from a smartwatch. This enables users to keep agents working on tasks like bug investigation or coordination while away from their computers.

Speed Improvements with Ultrafast

Amp is testing Ultrafast mode with GPT-6 Astra, which significantly reduces response times. This speed allows users to juggle fewer agents or gives agents more time to self-verify their work before returning results. Slack states that the combination of high quality and high speed is unprecedented, aiming to bring more pre- and post-coding work into the agent workflow.