Dispatches from the Amp team
RSSAmp now uses streamable HTTP transport for MCP servers by default with a fallback to Server-Sent Events.
That allows you to connect to your own MCP servers hosted on Cloudflare or use other remote servers such as https://learn.microsoft.com/api/mcp.
Amp now looks for AGENT.md files in subtrees, parent directories, and ~/.config/AGENT.md.
This lets you keep your top-level AGENT.md general and create more specific AGENT.md files in subtrees. They'll be included only when Amp works in that subtree.
You can also @-mention files in AGENT.md to include additional context.
There's something new in Amp's toolbox: a tool called oracle. Behind that tool is a read-only subagent powered by one of the most powerful models today: OpenAI's o3.
o3 is slightly slower than the model behind Amp's main agent, Sonnet 4. It's also slightly more expensive and less suited for day-to-day agentic coding. But it is impressively good at reviewing, at debugging, at analyzing, at figuring out what to do next.
That's why we made it available as a tool, so the main agent and the oracle can work together — hand in hand, or, well, the agentic equivalent: token by token, one writing code, the other analyzing and reviewing.
We consciously haven't pushed the oracle too hard in the system prompt, to avoid unnecessarily increasing costs for you or slowing you down. Instead, we rely on explicit prompting to get the main agent to consult the oracle.
Here are some prompts we used:
foobar and barfoo are used. Then I want you to work a lot with the oracle to figure out how we can refactor the duplication between them while keeping changes backwards compatible."
Amp now uses a different model when compacting or summarizing threads.
It's 4-6x faster, roughly 30x cheaper, and provides better summaries when you either compact a thread or create a new thread with a summary.
You can now queue messages that will only be sent to the agent once it's idle. To add a message to the queue, hold shift while submitting.
Amp now has subagents. Thorsten wrote a post about subagents, what they even are, and how they make us wonder whether everything is changing again.
Today, we're launching Amp Tab, our new in-editor completion engine, designed to anticipate your next actions and reduce the time spent manually writing code.
It uses a custom model that was trained to understand what you are trying to do next, based on your recent changes, your language server's diagnostics, and what we call semantic context.
Amp Tab can suggest regular single or multi-line edits to change entire code blocks, next to your cursor or farther away, somewhere else in your current document.
We're releasing Amp Tab now as a research preview, free to use for all Amp users. Enable it by adding the following to your VS Code settings:
{
"amp.tab.enabled": true
}
Amp now has improved support for multi-root workspaces:
Amp now provides a comprehensive Software Bill of Materials (SBOM) that catalogs all open source dependencies used. The SBOM follows the CycloneDX 1.6 specification and is compatible with security scanning tools and vulnerability databases. It's automatically generated during our build process and updated with each release.
What's included:
Access the Bill of Materials at /bom.json or learn more in the Amp Security Reference.
Amp now identifies secrets and redacts them with markers like [REDACTED:aws-access-key-id], so they are not exposed to the LLM, other tools, or ampcode.com. See Amp Security Reference for details.
Quinn and Thorsten discuss Claude 4, sub-agents, background agents, and they share "hot tips" for agentic coding.
The TODO feature that lets the agent manage its own list of tasks is now enabled by default.
We have made multiple changes to how we run terminal commands in Amp:
Environment — Commands now run by default in the integrated VS Code terminal inheriting your shell environment including Python virtual environments and direnv configuration.
Interactivity — Use the new "View in Terminal" button to interact with commands that require stdin, or the new "Detach" button to let the command continue running in the background.
Output — Progress bars with ANSI escape codes now render beautifully in the UI without eating up your context window, only the final output is presented to the model.
We've been using and tuning Amp with Claude 4 internally for weeks. Now that Anthropic has released it, we've switched Amp's primary model to be Claude Sonnet 4.
Amp is already designed to get the most power out of Claude 4. In our testing, we've found this new model to be:
To start using Claude 4 in Amp, update to the latest release of Amp's VS Code extension or CLI. No settings changes are needed; see why we don't make the user switch models.
You can now refer to image files in the Amp CLI. Paste the file path, drag a file into the terminal, or use file mentions by pressing @ to fuzzy-find the file. Images are useful for prompting with screenshots, designs, and more.
Last week we opened up Amp to the world, and now it's about to hit 500,000 messages sent.
We've been busy fixing bugs and making the whole experience of using Amp as smooth, solid, and fast as possible.
Please keep the feedback coming, on Discord, on X (1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 …), or to amp-devs@sourcegraph.com.
Thank you and happy coding!
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:
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.