Key Signals

• Cursor ships Composer 2, a frontier-level proprietary coding model that undercuts incumbents on price. Composer 2 scores 61.7% on Terminal-Bench 2.0 — surpassing Anthropic's Opus 4.6 (58.0%) — while pricing at $0.50/M input and $2.50/M output tokens, a fraction of Opus ($5/$25) or GPT-5.4 ($2.50/$15). The key technical innovation is "self-summarization," a compaction-in-the-loop RL technique that reduces context compression errors by 50%, enabling the model to maintain coherence across hundreds of actions in long-horizon coding tasks. This is the third Composer generation in five months, signaling Cursor's rapid trajectory from IDE wrapper to model developer. [1][2]

• OpenAI acquires developer tooling startup Astral to accelerate Codex growth. Astral, the company behind popular Python tools like Ruff and uv, will join OpenAI to power the next generation of Python developer tools within Codex. With Codex now serving over 2 million weekly users and experiencing three-fold user growth since January, this acquisition signals OpenAI's strategy of vertically integrating developer tooling to deepen its coding assistant ecosystem. [3][18]

• GitHub Copilot coding agent receives significant performance and observability upgrades. Copilot coding agent now starts work 50% faster, directly shortening the feedback loop for developers using it to create and iterate on pull requests. Alongside this, new session log improvements provide visibility into built-in setup steps, custom environment configuration, and subagent delegation — reflecting the growing complexity of agent workflows where orchestrating subagents is becoming a first-class concern. [4][5]

• Claude Opus 4.6 discovered 22 security vulnerabilities in Firefox in just two weeks, with 14 rated high-severity. This represents nearly 20% of all high-severity Firefox bugs patched throughout 2025, and Claude also wrote working exploits for two of the vulnerabilities. Mozilla validated the findings and shipped fixes in Firefox 148. Anthropic warns the gap between AI vulnerability discovery and exploitation capabilities is closing rapidly, underscoring the urgency for defenders to integrate AI-assisted security analysis into their workflows. [6]

• The AI industry is pivoting from browser agents to coding/terminal agents as the primary agentic paradigm. Google is restructuring its Project Mariner browser agent team, folding computer-use capabilities into its broader agent strategy including Gemini Agent. Terminal-based coding agents like Claude Code and OpenClaw have proven 10–100x more efficient than screenshot-based browser agents, fundamentally reshaping where AI labs are investing. OpenAI wants Codex to power general-purpose agents inside ChatGPT, while Anthropic has already launched Claude Cowork as a non-terminal offshoot. [7]

• Enterprise AI adoption data reveals a widening chasm between top adopters and laggards. Jellyfish's study of 700+ companies, 200K engineers, and 20M pull requests shows that 64% of companies now generate a majority of their code with AI assistance, and top-quartile adopters have doubled PR throughput. However, fully autonomous agent activity remains low but is growing exponentially, and the 90th-percentile companies have increased AI tool adoption ~7x year-over-year while the bottom quartile remains near zero. [8]

• MCP is driving a fundamental rethinking of enterprise API strategy. Morgan Stanley presented at QCon London how MCP has gone from obscurity to industry standard in 18 months, forcing the firm to redesign its five-year API program. Scaling beyond a handful of MCP tools creates disambiguation problems that confuse agents and waste tokens, pushing the industry toward specialized MCP gateways with business-context awareness — a shift away from traditional "dumb pipes" API gateway philosophy. [15]

AI Coding News

• OpenAI published research on monitoring internal coding agents for misalignment using chain-of-thought analysis. The work focuses on studying misalignment patterns in real-world coding agent deployments and detecting emerging risks to strengthen AI safety safeguards. This research arrives at a critical time, as coding agents become increasingly autonomous and capable of executing complex multi-step workflows with minimal human oversight. [14]

• AI coding models face a structural "staleness" problem that generic benchmarks fail to capture. At QCon London 2026, Jeff Smith argued that AI models trained on stale public repository snapshots consistently fail to follow the unwritten architectural and procedural rules that govern individual codebases. PR acceptance rates have actually declined even as AI-generated contribution volume has surged. Smith proposed "repository fingerprinting" — systematically extracting and documenting repository-specific constraints — to bridge the gap between generic model capabilities and production-ready code generation. [16]

• The AI industry shift toward coding agents is reshaping competitive dynamics across major AI labs. Google is restructuring Project Mariner, its Chrome browser agent initiative, as momentum shifts decisively toward terminal-based coding agents. Browser agent adoption has been disappointing — Perplexity's Comet reached only 2.8M weekly users and OpenAI's ChatGPT Agent fell below 1M — while coding agents have emerged as the more reliable and efficient paradigm. AI labs are now competing to extend coding agents beyond developer use cases into general-purpose assistants. [7]

• A practical tutorial demonstrates using Claude Code with Marp for AI-assisted presentation creation. The workflow uses a Claude Code skill with a /create-marp-deck slash command to interview users about their presentation goals, generate structured Marp Markdown drafts, and iteratively refine them through conversation or direct editing. The skill is compatible with Claude Code, Cursor, GitHub Copilot, and other AI agents, illustrating how agent skills are becoming portable across platforms. [17]

Feature Update

• Cursor Composer 2 launches as a frontier-level coding model with industry-leading cost efficiency. Priced at $0.50/M input and $2.50/M output tokens (with a faster variant at $1.50/$7.50), Composer 2 achieves 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual — substantial improvements over Composer 1.5 (44.2, 47.9, 65.9). The gains stem from Cursor's first continued pretraining run combined with "self-summarization," a compaction-in-the-loop RL technique that trains the model to compress its own context to ~1,000 tokens while preserving critical information, reducing compaction errors by 50%. [1][2]

• GitHub Copilot coding agent now starts work 50% faster. The optimization shortens the startup time for Copilot's cloud-based development environment, accelerating pull request creation and iteration cycles when using @copilot mentions or the Agents tab. This is particularly impactful for developers who use Copilot coding agent as part of their daily review-and-iterate workflow. [4]

• GitHub improves visibility into Copilot coding agent sessions with three key enhancements. Session logs now show built-in setup steps, custom setup step output from copilot-setup-steps.yml, and collapsible subagent activity panels with real-time status. These observability improvements address a growing need as coding agent workflows become more complex and involve multi-agent orchestration. [5]

• Copilot SDK v0.1.33-preview.3 adds OpenTelemetry tracing, blob attachments, and CJS compatibility. All four SDK languages now support distributed tracing with W3C trace context propagation across session operations. New features include a blob attachment type for inline base64 data, a reasoningEffort parameter for model switching, a skipPermission flag for low-risk tools, and Node.js CommonJS builds fixing VS Code extension compatibility. [10]

• Claude Code v2.1.80 ships with MCP channel previews, parallel tool fix, and memory optimizations. The release adds --channels allowing MCP servers to push messages into sessions, fixes --resume dropping parallel tool results, and reduces startup memory usage by ~80 MB on 250K-file repositories. Other additions include rate_limits statusline fields for Claude.ai usage tracking, inline plugin marketplace declaration via settings.json, and effort frontmatter for skills and slash commands. [11]

• Copilot CLI v1.0.9 resolves SSH disconnect noise and adds gitignored file search. Spurious I/O error messages no longer appear during SSH disconnects or terminal closes. A new include_gitignored config option enables @ file search to include gitignored files, CJK text copying on WSL is fixed, and marketplace/plugin installs from shortened URLs now work correctly. [9]

• OpenAI Codex 0.116.0 adds device-code sign-in, plugin installation prompts, and a user prompt hook. The app-server TUI now supports ChatGPT device-code onboarding and token refresh. Plugin setup is smoother with installation elicitation, a configured suggestion allowlist, and remote sync for install/uninstall state. A new userpromptsubmit hook enables prompts to be blocked or augmented before execution. Realtime sessions now start with recent thread context and are less prone to audio self-interruption. [12]

• Gemini CLI v0.35.0-preview.2 ships as a patch release. This minor release cherry-picks a targeted fix onto the v0.35.0-preview.1 branch. The previous nightly (v0.36.0) included more substantial changes such as sandbox integration for process-spawning tools, subagent support, and a disableAlwaysAllow security setting. [13]