Amp Expands Agents to Work Before and After Coding

OpenAI on October 14 published a case study on how coding agent company Amp uses GPT-6 Astra. Amp builds an agent that turns developers' ideas into working software. According to OpenAI, OpenAI models account for the majority of tokens used in Amp. The core of the case study is that agents are moving beyond writing code into a stage where they handle everything from deciding what to build to documentation, design and post-launch marketing.

Amp operates by choosing and recommending whichever models work best for its users. Quinn Slack, Amp's chief executive officer (CEO), cited transparency as a reason for choosing OpenAI models. "OpenAI is very transparent, and when there are safeguards applied, users know and they understand," he said. "You can do all of your work in Amp with OpenAI models and be really happy." He added that users value how consistently OpenAI models keep getting better.

The more work people hand to agents, the more token costs become a constraint. According to Slack, some Amp users consume 10 billion tokens a week. He believes that watching every increment of usage holds people back from finding out what an agent is capable of. That is why Amp launched its Sign in with ChatGPT integration, which gives users access to more tokens and to OpenAI models including GPT-6 Astra. "Sign in with ChatGPT means that people can have that feeling where they're not counting every single token," Slack said. "And once you actually let the agent go and have a longer leash, then you realize just how much it can do."

Amp's pricing structure has shifted in the same direction. In a September 13 announcement, Amp said it is now free to use when users bring their own compute and model subscriptions or API keys. A monthly plan is no longer required to use a ChatGPT subscription with Amp, and bring-your-own-key token fees and limits were dropped for everyone except the Enterprise tier. Amp instead charges for orbs, the remote computers where agents run independently and in parallel, while running agents on runners on users' own computers is free. In the same announcement, Amp said it is formally extending its zero and minimal data retention policy to all users, not just Enterprise customers, and on September 30 it opened nine additional model routing options, including OpenRouter and Amazon Bedrock, to every tier.

The scope of what agents handle has widened too. Amp's 20-person team uses agents for marketing as well. "We don't have someone devoted to it, but we have an agent devoted to it," Slack said. That agent works on improving the experience for new users and helping more people discover Amp through search and other channels. In that work, the team uses GPT-6 Astra to produce documentation and design pages. "We have a very unique design style and writing style. We're very picky about this, but we've been really happy with how GPT-6 Astra does," Slack said. He also gave GPT-6 Astra the plans for his house, used it to build a 3D model in Blender and then turned that model into games with his children, adding that this was not possible just a couple of months ago.

Puck, Amp's conversational assistant, extends the agent into voice. Puck is built into Amp to launch and manage other agents: it can start an agent to investigate a bug, coordinate several agents or find a previous conversation. With Calling Puck, powered by GPT-Live-1, users can talk to Puck in real time, even from a watch. On October 13, Amp opened Puck as an MCP server so it can be called from ChatGPT, Claude and other agents. According to Amp's announcement, which METAL reviewed, Puck MCP offers two tools: one for talking to Puck and one for showing running orbs.

Speed is the next variable. Amp has been testing Ultrafast mode with GPT-6 Astra, and on September 29 it introduced Plaid speed, which uses OpenAI's Ultrafast speed tier. According to Amp's announcement, Plaid runs inference requests up to six times faster, at six times the cost per token. For now it works only with Amp-provided inference, because OpenAI does not yet support Ultrafast for linked ChatGPT subscriptions. METAL previously reported that OpenAI launched its Ultrafast speed tier. "People have not seen a model that's this good, that's this fast," Slack said. He expects faster responses to reduce how often people juggle several agent tasks while waiting, and to give agents more time to check their own work before returning a result.

From a content marketer's perspective, the notable point in this case is that a 20-person company assigned its marketing role to an agent rather than a person. The division of labor gives measurable work such as search traffic and the new-user experience to the agent, while the team keeps control of the standards for writing and design. Amp's move to lower the barrier with a free tier and the removal of bring-your-own-key fees stems from the same calculation: getting users to hand bigger jobs to agents. Amp's next step is to bring more of the work before and after coding into the agent. "There's this unknown of what humans will create, and that's what we all get to see," Slack said. "It's cool to be playing a part in that."