Key Signals

• The US government designated Anthropic a "supply-chain risk," sending shock waves through the AI coding ecosystem. Defense Secretary Pete Hegseth directed that "no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic," after weeks of failed negotiations over military AI use restrictions. This unprecedented designation — historically reserved for foreign adversaries — could force defense contractors like Amazon, Microsoft, Google, and Nvidia to re-evaluate their relationships with Anthropic, potentially disrupting access to Claude Code and Claude Gov for developers working in defense-adjacent sectors. Legal experts say the full scope remains unclear, but the chilling effect on AI companies' willingness to negotiate government terms is immediate. [1][8]

• OpenAI signed a deal to deploy models in the Pentagon's classified network, positioning Codex as the de facto government AI coding tool. CEO Sam Altman announced the agreement includes prohibitions on domestic mass surveillance and requires "human responsibility for the use of force, including for autonomous weapon systems." However, analysts note that OpenAI's language commits to less than what Anthropic demanded — Anthropic wanted explicit human-in-the-loop before any lethal action, while OpenAI's framing allows for accountability after the fact. Over 60 OpenAI employees and 300 Google employees had signed an open letter supporting Anthropic's position earlier that week. [2][9]

• Claude Code v2.1.63 shipped with new slash commands, HTTP hooks, and a wave of memory leak fixes critical for long-running agentic sessions. The release adds /simplify and /batch bundled slash commands, HTTP hooks that can POST JSON to a URL and receive JSON responses, and project configs shared across git worktrees. Crucially, the release addresses at least nine distinct memory leak and listener leak issues that affected long-running sessions — including leaks in MCP server caches, git root detection, JSON parsing caches, and WebSocket transports — making Claude Code significantly more stable for extended agentic workflows. [3]

• Andrej Karpathy declared "vibe coding" obsolete one year after coining the term, proposing "agentic engineering" as the replacement paradigm. In a viral post (3M+ views), Karpathy described giving an agent a single English-language prompt to set up an entire video analysis dashboard — the agent ran for 30 minutes, hit multiple issues, researched solutions online, debugged, and delivered the finished product. This shift aligns with growing consensus that coding agents crossed a usability threshold around December 2025, and that the new distribution channel for software is agents calling CLIs, MCP servers, and docs programmatically. [4]

• Anthropic's Agent Skills repository reached 76K+ GitHub stars and is becoming a cross-tool standard adopted by Microsoft, Cursor, Goose, and OpenAI. The Skills spec — now at v0.9 under the Linux Foundation-backed Agentic AI Foundation — provides folders of instructions, scripts, and resources that teach AI agents repeatable workflows. Microsoft adopted it inside VS Code and GitHub, while OpenAI quietly adopted a structurally identical architecture in ChatGPT and Codex, suggesting convergence on a shared agent capability layer. [4]

• Gemini CLI's nightly build introduced policy engine support for extensions and a G1 AI credits billing flow. The v0.33.0-nightly.20260228 release includes over 45 changes, with notable features like contiguous parallel admission for sub-agent tools, a temporary flag to disable workspace policies, improved planning workflows encouraging multi-select with option descriptions, and fixes preventing sub-agent tool calls from leaking into the UI. These changes signal Google's push toward enterprise-grade agent governance in Gemini CLI. [5]

AI Coding News

• Anthropic's Claude app surged to #2 in the US App Store amid the Pentagon standoff, trailing only ChatGPT. According to SensorTower data, Claude was outside the top 100 at the end of January, spent most of February in the top 20, and climbed from sixth on Wednesday to second by Saturday as the government dispute drew massive public attention. Google Gemini sits at #3, marking the first time the three major AI chatbots have simultaneously dominated the App Store's top spots. [6]

• Perplexity launched "Computer," a general-purpose digital worker that deploys subagents across Gmail, Slack, Notion, and hundreds of other apps. Unlike OpenClaw, which requires local setup, Perplexity Computer runs in a browser and can operate for hours or even months on a single task. The launch represents a significant shift from AI agents as developer tools to AI agents as mainstream productivity interfaces, with the tagline "Chat answers. Agents do tasks. Computer works." [4]

• Cloudflare's "Markdown for Agents" feature reduces token consumption by 80% when AI agents browse the web. When an agent requests a webpage, Cloudflare's edge converts HTML to clean markdown before returning it — Cloudflare's own blog post dropped from 16,180 tokens in HTML to 3,150 in Markdown. Claude Code and OpenCode already send the Accept headers that trigger this conversion, and the feature is in beta for Pro, Business, and Enterprise customers at no additional cost, making every Cloudflare-protected website instantly more agent-accessible. [4]

• India disrupted access to Supabase, one of the most popular backend-as-a-service platforms among developers. India is one of Supabase's biggest markets, and patchy access following a government blocking order could affect thousands of developers relying on the platform for database hosting, authentication, and real-time APIs. The disruption highlights increasing geopolitical fragmentation of the developer tool ecosystem. [10]

• A Lobsters-trending essay argued that segregating workflows between AI agents and humans — via files like AGENTS.md or flags like AI=true — is an anti-pattern. The piece by Vladimir Keleshev makes the case that documentation placed in copilot-instructions.md is equally valuable for humans, tools exposed as MCP servers should default to CLI APIs, and developers should design for unified human-and-agent interfaces rather than maintaining parallel workflows. The essay resonated with the developer community's growing unease about tool fragmentation in the agentic era. [11]

Feature Update

• Claude Code v2.1.63 adds /simplify, /batch bundled slash commands and HTTP hooks for programmatic integrations. HTTP hooks can POST JSON to a URL and receive JSON instead of running shell commands, enabling richer CI/CD and webhook integrations. Project configs and auto memory are now shared across git worktrees of the same repository, and an ENABLE_CLAUDEAI_MCP_SERVERS=false env var lets users opt out of claude.ai MCP server availability. The /model command now shows the currently active model, and /copy adds an "Always copy full response" option. VSCode gets session rename/remove actions and a fix for remote sessions not appearing in history. Nine memory leak and listener leak fixes improve stability for long-running sessions. [3]

• OpenAI Codex released 0.107.0-alpha.8 and 0.107.0-alpha.7, continuing rapid iteration on the Rust-based CLI. These are incremental alpha builds following the major 0.106.0 stable release which introduced a direct macOS/Linux install script, expanded app-server v2 thread API with realtime endpoints, promoted js_repl to experimental, made 5.3-codex visible in the CLI model list, and added diff-based memory forgetting with usage-aware selection. The alpha cadence — eight releases in two days — signals active development ahead of the next stable release. [7]

• Gemini CLI v0.33.0-nightly.20260228 ships policy engine for extensions and parallel sub-agent tool admission. Key additions include support for policy engine in extensions, G1 AI credits overage flow with billing telemetry, improved planning workflows that encourage multi-select with descriptions of options, and contiguous parallel admission for Kind.Agent tools. Fixes address sub-agent tool calls leaking into the UI, shell autocomplete polish, intrusive MCP error deduplication, and alternate buffer mode dialog scrolling. The release also adds low/full CLI error verbosity modes for cleaner UI output. [5]