Amp Code: Next generation AI Coding (AI column)
I saw an interesting one. report Beyang Liu, CTO Sourcegraph, about the new Amp Code editor, in which the author says that this is not just “another Copilot”, but an attempt to fundamentally change the way we interact with AI in development. In short: according to the author, we are moving from the phase “AI writes text” to the phase “AI closes the development cycle”. Beyang Liu graduated from Stanford and worked at Google and Palantir. Known as the Creator of Sourcegraph (code-searcher) Cody (One of the first AI assistants with codebase context). He believes that the main thing in AI coding is not the generation of tokens, but access to the code knowledge graph and runtime environment.
The main points of the report are as follows:
1Death of the "Copilot Paradigm" Traditional AI assistants (GitHub Copilot, early Cody) It works like a "smart car kit." They predict the next token, but don’t know if the code works. Beyang calls it "Fire and Forget": AI gave out the code, and it's up to you to solve compilation errors.
2️⃣ Agentic Loop Amp Code builds on the cycle OODA (Observe-Orient-Decide-Act)
- AI writes code.
- You run the linter/compiler/tests
- Sees a mistake. (for example,
TypeError) - Corrects it without human intervention.
- Repeat until it works.
3Context is not just text. Just stick it in. 100 files in the context window (even on 1M tokens) - not enough. Amp uses the LSP (Language Server Protocol) and real data from runtime to understand the structure of the project as IDE understands it, not just as a set of characters.
4Review Agent Mode Amp has a separate agent reviewer built in. Before applying changes, he conducts Code Review: looking for bugs, checking style and security, simulating the PR review process in the team.
🚀 What does that mean for development?
- Skillset shift: We're moving from "quick code set" to managing agents. Your job is to clearly state your intention. (Intent) architecture and implementation (Implementation) And the tule takes over.
- Less Context Switching: You don’t need to switch between the editor and the terminal to check if the code that the AI has issued is working. The agent does it in the background. Unix-way: Beyang emphasizes that Amp is available as both a VS Code extension and a CLI tool. It’s a return to the roots: powerful tools that can be scripted and built into piplins.
In the Amp report and documentation, Beyang draws on the following concepts and materials: 1. Agentic Workflows & Scaling Laws The author refers to the fact that the quality of the code does not grow linearly from the size of the model, but jumps when using agentic loops. This is confirmed by the results of the SWE-bench benchmark, where agents who can run code radically bypass simple LLMs. More about the concept readable Andrew Ng 2. Sourcegraph’s "Big Code" Intelligence Amp database is a graph code analysis technology (SCIP)This is something that Sourcegraph has been working on for years. 3. LSP as a source of truth LLMs need structured data from a compiler, not just text. This is a reference to the Language Server Protocol, which was developed by Microsoft for its code editor VS Code, but has become an open standard and is now actively developing in conjunction with Red Hat and Codenvy, and the project itself is posted on the Internet. GitHubThis allows you to use it in different editors and for many programming languages.
#AI #Software #Engineering #Architecture #Agents #ML #SystemDesign