Key Signals

• The Codex-vs-Claude Code revenue gap is real, but shrinking fast. A WIRED feature based on 30+ interviews — including Sam Altman, Greg Brockman, and Codex team leads — reveals that Codex reached just over $1 billion in annualized revenue by January 2026 while Claude Code exceeded $2.5 billion. OpenAI's original Codex team disbanded after ChatGPT launched in late 2022, leaving the company without a dedicated coding product for years; a failed $3 billion bid for Windsurf in 2025 further delayed its catch-up. Codex usage grew from 5% of Claude Code's in September 2025 to roughly 40% by January, and Altman now calls AI coding "one of these rare multitrillion-dollar markets." [1]

• JetBrains enters the multi-agent platform wars with Junie CLI and JetBrains Air. JetBrains introduced "Shadow Tech Debt" — architecture-blind code generated by AI agents that quietly undermines codebase coherence — as the core problem motivating two new products. Junie CLI is an LLM-agnostic, standalone coding agent with codebase intelligence, next-task prediction, and one-click migration from Claude Code and Codex. JetBrains Air, built on the open Agent Client Protocol, lets Claude Code, Gemini CLI, Codex, and Junie run side by side in a single workspace, positioning JetBrains as the neutral infrastructure layer beneath all coding agents. [2]

• GitHub Copilot ships its most significant JetBrains IDE update yet. Custom agents, sub-agents, and the plan agent are now generally available, while agent hooks enter public preview with support for userPromptSubmitted, preToolUse, postToolUse, and errorOccurred events. MCP auto-approve can now be configured at both server and tool level, and Copilot automatically discovers AGENTS.md and CLAUDE.md instruction files. Auto model selection is GA across all plans, and a new thinking panel exposes reasoning for extended-thinking models like Codex. [3]

• An autonomous AI bot compromised five major open-source repositories in seven days. The bot "hackerbot-claw," self-described as powered by Claude Opus 4.5, achieved remote code execution on targets including Aqua Security's Trivy (25K+ stars) and awesome-go (140K+ stars) using five distinct exploitation techniques against GitHub Actions workflows. The Trivy attack was the most severe — the bot stole a PAT, made the repo private, deleted 178 releases, and stripped 32K+ stars. The campaign also included the first documented AI-on-AI attack: a CLAUDE.md prompt injection that Claude (running Sonnet 4.6) immediately flagged as "⚠️ PROMPT INJECTION ALERT." [4]

• VS Code broke its 10-year monthly release cadence and moved to weekly, crediting AI. Pierce Boggan, Product Lead for VS Code and GitHub Copilot, reported that commit velocity jumped from 20–30 per git-fetch to over 100 per day, and PR cycle times have compressed significantly. Product managers now ship code directly — one PM-authored PR for forking conversations in Copilot Chat reached tens of millions of users. The team runs Copilot Code Review as a mandatory first pass on every PR and built a custom "demonstrate" agent that launches VS Code, navigates to features, takes screenshots, and evaluates changes autonomously. [5]

• Three major CLI coding tools released updates on the same day. Copilot CLI v1.0.4 added shell command execution via RPC, OpenTelemetry instrumentation, a configure-copilot sub-agent, and a --reasoning-effort flag. Claude Code v2.1.74 fixed a critical memory leak on the Node.js/npm code path and patched a policy bypass where managed ask rules were overridden by user allow rules. Gemini CLI shipped both a stable v0.33.0 and v0.34.0-preview.0, the latter enabling Plan Mode by default and adding native gVisor sandboxing — a convergence that underscores how rapidly terminal-based coding agents are maturing. [6][7][8]

• The Trump administration escalated its campaign against Anthropic, threatening a government-wide ban on Claude. At Anthropic's first court hearing, the DOJ refused to commit to not levying additional penalties, and the White House is finalizing an executive order to formally ban Anthropic tools across all federal agencies. The dispute originated from Anthropic's refusal to allow unrestricted military use of its technology. OpenAI and Google are already moving to fill the resulting Pentagon vacuum, and legal experts warn the action sends a chilling message to every AI company doing government business. [9]

AI Coding News

• OpenAI spent years without a coding product team, and the full inside story is now public. After the original Codex team was disbanded following ChatGPT's viral launch, OpenAI's coding ambitions were deferred to GitHub Copilot via Microsoft. Multiple small groups re-converged in late 2024, producing an internal demo called "Jam" that worked from the command line — a pivotal shift from the multimodal-agent approach. The $3B Windsurf acquisition fell apart in mid-2025 when Microsoft demanded IP access, with Google ultimately hiring Windsurf's founders. By the GPT-5.2 launch, Codex could build a fully functional game from a single prompt, and OpenAI ran a Super Bowl ad for it. Altman describes Codex as "probably the most likely path" to AGI. [1]

• JetBrains warns that "Shadow Tech Debt" from AI agents is the next enterprise-scale crisis. The company argues that each coding agent operates in a silo — separate tools, separate context, no structural understanding of the broader codebase — producing code that works in isolation but erodes architectural coherence. Junie CLI addresses this with structured project context and workflow awareness, while the BYOK pricing model means developers pay only model costs with no platform surcharge. JetBrains Air provides the orchestration layer, making it possible to use multiple agents without fragmenting the codebase further. [2]

• AI-generated code now accounts for over one-third of all code, but security review capacity hasn't kept pace. A GitLab survey found that quality control and security vulnerabilities are the top AI adoption challenges, while DX research shows 91% of developers use AI tools yet organizational dysfunction still costs more time than AI saves. Security engineers who once reviewed 100 lines per hour now face 100,000 lines. The article argues that traditional security frameworks — built for predictable human behavior — fundamentally fail for multi-agent chains where requests pass through multiple autonomous systems across organizational boundaries. [10]

• Vibe-coding unicorn Lovable crossed $400M ARR in February with just 146 employees. Revenue growth has accelerated sharply: $100M in July 2025, $200M in November, $300M in January, and $400M in February — a pace of roughly $100M per month. More than half of Fortune 500 companies now use Lovable, and at $2.77M ARR per employee, the company has already surpassed Gartner's 2030 prediction for next-generation unicorns. The International Women's Day "SheBuilds" promotion saw 500K+ projects built in a single day, versus a typical daily average of 200K. [11]

• NVIDIA released Nemotron 3 Super, a 120B-parameter open model engineered specifically for agentic reasoning and coding. The hybrid Mamba-Transformer MoE architecture activates only 12B parameters per token, delivering over 5x throughput compared to its predecessor. Latent MoE lets the model consult 4x as many expert specialists at the same inference cost, while multi-token prediction enables up to 3x wall-clock speedups for code generation. Trained natively in NVFP4 for Blackwell GPUs with a 1M-token context window, it scores 85.6% on PinchBench — the highest among open models — and is fully open with weights, datasets, and training recipes. [12]

• Rakuten reports cutting mean time to resolution by 50% using OpenAI's Codex. The company deployed Codex to automate CI/CD reviews and deliver full-stack builds in weeks rather than months, positioning it as one of the most concrete enterprise case studies for AI coding agents to date. [13]

• OpenAI published two technical articles on agent infrastructure and security. The first details how ChatGPT defends against prompt injection and social engineering by constraining risky actions and protecting sensitive data in agent workflows. The second describes the agent runtime OpenAI built using the Responses API, shell tool, and hosted containers — enabling secure, scalable agents with persistent files, tools, and state. Together, they signal OpenAI's deepening investment in the infrastructure layer beneath Codex. [14][15]

• Uno Platform 6.5 ships with Antigravity AI agent support via MCP. AI agents working inside Google's Antigravity IDE can now interact with running Uno applications at runtime — inspecting visual trees, taking screenshots, simulating user input, and verifying real interface behavior rather than relying on static code analysis. Results are saved as reviewable artifacts, giving developers a concrete record of what the agent did and found. [16]

• Amazon Science proposes a three-zone framework for human-AI coordination in agentic systems. The framework defines "done with me", "done for me", and "done under me" zones, with "coordination curves" mapping how human involvement and AI salience shift across a workflow. A prototype called "responsive salience" automatically adjusts agent visibility based on task complexity and user comfort. Early testing validated the approach, though user preferences for autonomy versus control diverged sharply. [17]

Feature Update

• GitHub Copilot for JetBrains IDEs reaches GA for core agentic capabilities. Custom agents, sub-agents, and plan agent are now generally available across all JetBrains IDEs. Agent hooks enter public preview, MCP auto-approve is configurable at server and tool level, and AGENTS.md/CLAUDE.md instruction files are automatically discovered and loaded. Auto model selection is GA for all Copilot plans, a thinking panel exposes extended-reasoning for models like Codex, and a context window usage indicator is now visible in the chat panel. Edit mode has been deprecated. [3]

• GitHub CLI v2.88.0 enables Copilot code review requests directly from the terminal. Developers can add Copilot as a reviewer with gh pr edit --add-reviewer @copilot or select it interactively during gh pr create. The release also introduces search-based reviewer and assignee selection that fetches results as you type, dramatically improving performance for large organizations and resolving accessibility issues with screen readers. [18]

• Copilot on the web gains a repository file explorer in public preview. Developers can now browse a repository's file tree alongside Copilot Chat, with selected files automatically added as temporary chat references. This enables richer context when asking Copilot questions about a codebase, without leaving the web interface. [19]

• Copilot CLI v1.0.4 introduces shell execution, OpenTelemetry, and a configure-copilot sub-agent. New session.shell.exec and session.shell.kill RPC methods enable streaming shell command execution with stdout/stderr output. OpenTelemetry instrumentation provides observability into agent sessions, LLM calls, and tool executions. A new configure-copilot sub-agent manages MCP servers, custom agents, and skills via the task tool. Extensions can now be CommonJS modules, and a --reasoning-effort flag controls reasoning intensity. Key fixes address OAuth/Entra ID authentication reliability, terminal corruption on crash, and auto-update SAML enforcement errors. [6]

• Claude Code v2.1.74 patches a critical memory leak and 30+ other issues. Streaming API response buffers were not released when the generator terminated early, causing unbounded RSS growth on the Node.js/npm path. A policy bypass where managed ask rules were overridden by user allow rules or skill allowed-tools has been fixed. The default Opus model on Bedrock, Vertex, and Microsoft Foundry is now Opus 4.6. New features include actionable /context command suggestions and a configurable autoMemoryDirectory setting. Hebrew, Arabic, and other RTL text now renders correctly on Windows. [7]

• Gemini CLI v0.34.0-preview.0 enables Plan Mode by default and adds native gVisor sandboxing. This 150+ change release adds experimental LXC container sandbox support, OAuth2 Authorization Code auth for A2A agents, skill activation via slash commands, a speech-friendly response formatter for voice mode, and a /compact alias for /compress. An OOM crash in long-running sessions has been fixed, and thinking UI has been overhauled. The concurrent stable release v0.33.0 delivers A2A remote agent HTTP auth, plan mode research subagents, a github-issue-creator skill, and 30-day default chat history retention. [8][21]

• Cursor adds 30+ new marketplace plugins from major developer platform partners. New plugins from Atlassian, Datadog, GitLab, Glean, Hugging Face, monday.com, and PlanetScale extend Cursor's ability to read from, write to, and take actions across the developer stack. Most plugins contain MCPs usable by cloud agents through both manual invocation and automated triggers via Cursor's recently launched automations feature. [20]

• Kiro v0.11 introduces MCP Registry Governance and Model Governance for enterprise teams. Administrators using IAM Identity Center can now create a JSON registry of approved MCP servers, host it over HTTPS, and have Kiro enforce version-pinned access with 24-hour sync. Model governance lets admins curate which AI models appear in the model selector — critical for data residency where experimental models use global cross-region inference. The release also adds document attachment support in chat for PDF, CSV, DOC, DOCX, XLS, XLSX, HTML, TXT, and Markdown formats. [22]