News

As of today, Amp is available to everyone. No more waitlist.

Thorsten used the occasion of Amp now being available to everyone to write down how he uses it and how it changed programming for him.

Beyang interviews Thorsten and Quinn to unpack what has happened in the world of Amp in the last five weeks: how predictions played out, how working with agents shaped coding practices, and the evolution of AI coding tools from browser automation to model training.

You can now navigate the local history of messages in the Amp CLI. Use the arrow keys or Ctrl+P/Ctrl+N to navigate the adjacent messages in the history. PageUp and PageDown jumps directly, regardless of cursor position. Your current draft is available at the end of the history, and you can press Ctrl+C or Esc to cancel and go back to your draft.

We removed Isolated Mode, which let you use Amp with your own API keys for LLM inference, because it's not possible for it to meet our quality bar. The intent was to make it easier to use Amp in locked-down environments, but we hit many issues that made the experience bad and slowed us down:

1. Anthropic rate limit issues: Individuals can't easily get Anthropic API keys with high enough rate limits to actually use Amp. (Anthropic is understandably strict here unless you are an enterprise customer of theirs.)
2. LLM proxy and connectivity issues: Many people tried to use Isolated Mode with custom LLM proxies that weren't fully API-compatible with Anthropic's API, which led to hard-to-debug issues.
3. Gemini API parity issues: We found that Gemini 2.5 Pro only works well via Google Cloud Vertex AI (a more enterprise-y offering) rather than Google AI Studio, which is how most people would generate API keys, because of differences in how they handle thinking.

Even though we could work around these specific issues, more will arise in the future because tool-calling agents need to integrate more deeply into model capabilities and LLM APIs are getting more complex and differentiated.

We believe the best product is built by iterating fast at the model↔product frontier, and most devs and companies want the best coding agent more than they want a worse coding agent that satisfies other constraints.

If you were using Isolated Mode: When you update Amp in VS Code, you'll see a message informing you of this change and requiring you to disable Isolated Mode to continue. Your threads are preserved locally.

The Amp CLI now supports file mentions in interactive mode. Type @ followed by a pattern to fuzzy-search. Use Tab or Shift-Tab through the results, and hit Enter to confirm.

Amp now looks in the AGENT.md file at the root of your project for guidance on project structure, build & test steps, conventions, and avoiding common mistakes. Manual »

Amp will offer to generate this file by reading your project and other agents' files (.cursorrules, .cursor/rules, .windsurfrules, .clinerules, CLAUDE.md, and .github/copilot-instructions.md).

We chose AGENT.md as a naming standard to avoid the proliferation of agent-specific files in your repositories. We hope other agents will follow this convention.

An operator's guide to Amp

It’s not that hard to build a fully functioning, code-editing agent.

Our responses to some common feedback that we are intentionally not acting on.

What will AI do to open-source? What does it mean for GitHub? What does it mean for interviewing engineers?

What does this all mean for code search? How do you balance coding knowledge/skills/experience vs. letting the AI do it?

Is the magic with these agents that there are no token limits? Where do the agents fail? Do they need guidance? Where? How does one even price this? Isn't it too expensive?

Thorsten and Quinn start building Amp.