Why AMP’s Removal of Its VS Code Plugin Signals the End of Human‑AI Pair Programming

AMP announced on a February 2026 podcast that it will shut down its VS Code plugin and Cursor extension within 60 days, stating that the IDE sidebar model that has defined AI‑assisted coding for the past two years is no longer viable.

Deep Mode: When AI Learns to Think Deeply

Before 2025, mainstream models such as Claude 3.5 Sonnet were described as “smart but impatient,” excelling at Smart Mode where a question receives an immediate code snippet or fix. With the release of GPT‑5.2 Codex, AMP introduced Deep Mode . The model accepts a vague, high‑level goal (e.g., “refactor the entire authentication module for a new security protocol”), then runs autonomously for 45–60 minutes, searching documentation, trying solutions, self‑correcting errors, and executing tests until a final result is delivered. The sidebar cannot accommodate such long‑running interactions.

“You look at the code, the AI looks at you in the sidebar, back‑and‑forth… this pattern is not the future. For the 1 % who want to live in the future, the sidebar is a shackles rather than a boost.”

Conclusion 1: When AI capabilities shift from second‑level completions to hour‑level tasks, the AI must detach from the IDE and run as an independent worker rather than an editor‑bound assistant. Reference: https://mp.weixin.qq.com/s?__biz=MzIyNzM0MDk0Mg==∣=2247504858&idx=1&sn=031ff6f85643affc2887e97e535175fb&scene=21#wechat_redirect

Agent DX > Human DX

To improve internal efficiency, the AMP team rewrote svelte-check in Zig and renamed it zvelt-check. The new tool runs faster and produces more structured logs for parsing, but it degrades VS Code’s native Svelte support, reducing error highlighting and other human‑focused diagnostics. Some developers, especially those using NeoVim, expressed indifference, saying they only care about Agent speed.

Conclusion 2: Future toolchains will prioritize AI optimization; a tool that is unfriendly to humans but friendly to agents will still be adopted, enabling up to a ten‑fold productivity boost.

Software Melting: From SaaS to Text

The interview introduced the concept of “The Melting of Software,” suggesting that traditional GUI‑based applications will give way to pure services.

Case A – Ryan Florence’s fitness coach: Ryan uses ChatGPT’s voice mode to receive real‑time workout instructions without any app or UI. The interaction is entirely conversational, showing software disappearing into a service.

Case B – Todoist replacement: AMP’s founder considered automating Todoist, then realized a plain‑text file managed by a Skill could handle a 15‑item shopping list, eliminating the need for a complex SaaS UI.

Conclusion 3: Software is devolving into APIs and data; the interaction layer is being taken over by Agents.

Skills: A New Abstraction Layer

AMP demonstrated extensive internal use of Skills, which encode expert knowledge for Agents. Examples include:

Tmux Skill – teaches an Agent how to correctly use Tmux and terminate processes, even handling nuanced steps like pressing Ctrl‑C twice.

Google Cloud Skill – equips an Agent with the ability to invoke the Google Cloud CLI.

BigQuery Skill – the Agent writes SQL, queries BigQuery, and returns results on demand.

Skills are “codified experience.” After an Agent solves a problem, the solution is distilled into a Skill so that the same or other Agents can repeat the process without error, far more efficiently than re‑prompting in a chat window.

Organizational Philosophy

AMP’s decision to cut the VS Code plugin reflects a philosophy of rapid iteration: technology cycles every three months, and clinging to legacy habits is seen as a trap. The team cited examples such as GitHub Copilot’s waning relevance after Cursor, Claude Code, and the upcoming OpenClaw agents.

AMP’s CEO warned that preserving features for user habit would make the company a “best‑of‑the‑old” laggard, emphasizing the need to win customers anew every quarter.

Final Outlook

For the 1 % of developers willing to adopt “Factory Mode,” productivity will shift from linear to exponential growth. The future developer will define specifications, command Agents from the terminal, and manage Agent clusters rather than writing code directly. The sidebar is dead; the software factory lives on.