Amp is moving beyond editor sidebar “agents,” arguing newer coding models don’t need heavy wrappers. It’s retiring its VS Code and Cursor extensions on March 5 at 8pm PT, while keeping the Amp CLI as the main interface—for now.
LangChain says it boosted its deepagents-cli score on Terminal Bench 2.0 from 52.8% to 66.5%—without changing the underlying model. The gains came from “harness engineering”: tighter prompts, smarter middleware, trace-driven iteration, and a better verify loop.
Vercel's evaluation shows embedding a compressed docs index in AGENTS.md let agents achieve 100% pass rates on Next.js 16-specific coding tests, outperforming on-demand skills. Passive context removed invocation and sequencing fragility while an 8KB index kept prompt size manageable.
A recent piece by Addy Osmani takes a closer look at developers handing most code to AI agents and the hidden cost of “comprehension debt.” He warns faster generation can create a new review bottleneck and outlines when agent-first workflows make sense.
An engineering write-up explains how OpenAI's Codex CLI runs an agent loop: it builds structured prompts, streams Responses API output, executes local tools (like shell commands), appends results, and re-queries the model until completion. It also outlines prompt caching, stateless requests, and …