The Agentic IDE Wars: Cursor, Claude Code, and OpenAI Codex Updates

What shipped

The landscape for AI-assisted coding is shifting from simple chat windows to complex agentic workflows. This week, we saw significant movement in how these tools handle multi-tasking, enterprise-grade governance, and deep integration into local development environments.

Cursor

Cursor continues to solidify its lead as the primary environment for teams by focusing on visibility and administrative control. These updates provide the structure needed to manage complex projects where multiple agents and subagents operate simultaneously.

• •PR Review, Build in Parallel, and Split PRs _(3.3)_: PRs (pull requests, the team-review step before code merges) get a new review interface with inline threads and commit history. 'Build in Parallel' runs independent tasks at the same time using async subagents (smaller AI helpers running underneath). A quick action splits one big change into separate PRs with automatic dependency detection.
• •Context Usage Breakdown _(3.3)_: Detailed token-usage breakdown (tokens are billable text chunks the model reads and writes) per agent across rules, skills, MCPs (Model Context Protocol servers, how agents plug into external tools), and subagents. Lets you see which part of your setup is bloating context and trim it.
• •Model controls, spend management, usage analytics Enterprise admins get model blocklists by provider, soft spend limits with 50 / 80 / 100 percent alerts, and usage analytics by user and product surface (Cloud Agents, Security Review).
• •Team Marketplace Updates Build team plugin marketplaces from settings, no git repository (a versioned code store) required. Three distribution modes: Default Off (opt-in), Default On (opt-out), and Required. Plugins bundle MCP servers, skills, subagents, rules, and hooks.

View Cursor's full changelog on cursor.com ↗

Claude Code

Anthropic is refining the experience for terminal-heavy users by addressing stability and environment-specific constraints. The focus here is on ensuring that the agent behaves predictably when running in restricted corporate networks or complex local setups.

• •Worktree improvements, sandbox enhancements, bug fixes _(2.1.133)_: A worktree (a separate working copy of your codebase pinned to a specific branch) can now branch from origin/ (a fresh starting point) or local HEAD (your current state). Fixed parallel sessions dead-ending at HTTP 401 (the unauthorised error code) from a refresh-token race condition. Fixed corporate proxy variables (HTTP_PROXY, HTTPS_PROXY) and mTLS (mutual TLS authentication, where both sides verify each other's identity) being ignored during the full MCP OAuth flow.
• •Environment variables, terminal rendering, session fixes _(2.1.132)_: New environment variables CLAUDE_CODE_SESSION_ID and CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 (opts your terminal out of taking over the whole screen). Fixed external SIGINT (signal interrupt, the Ctrl+C / IDE-stop-button signal) not running a graceful shutdown. Fixed mouse-wheel scroll being too fast in Cursor and VS Code (caused by an upstream bug in xterm.js, the terminal-rendering library).
• •Critical authentication hotfix _(2.1.131)_: Fixed the VS Code extension failing to activate on Windows due to a hardcoded build path. Fixed the Mantle endpoint authentication failing because the x-api-key header was missing.
• •Plugin URLs, effort level awareness, bug fixes _(2.1.129)_: New --plugin-url flag fetches plugin .zip archives directly from a URL for the current session. Hooks (small scripts that fire at lifecycle events) now receive the active effort level via the effort.level JSON field and the $CLAUDE_EFFORT environment variable.

View Claude Code's full changelog on code.claude.com ↗

OpenAI Codex

Codex is expanding its reach beyond the local terminal and into the browser, aiming to capture the workflow of developers who live in web-based environments. The introduction of modal editing and cross-tab awareness signals a push toward more sophisticated input methods.

• •Codex CLI 0.129.0 _(CLI 0.129.0)_: Modal Vim editing (Vim is a keyboard-driven text editor; modal means it has separate modes for typing vs navigating) lands in the TUI (text-based user interface, the in-terminal UI) composer with configurable keymaps. Redesigned resume / fork picker with raw scrollback mode and workspace-aware /diff routing. Plugin management adds workspace sharing, access controls, marketplace removal, admin-disabled status handling, and experimental goals that stay paused across sessions.
• •Codex for Chrome Extension A new Chrome extension lets Codex work in parallel across multiple browser tabs without taking over the browser. Users keep full control over which websites Codex can access.

View OpenAI Codex's full changelog on developers.openai.com ↗

What this means for you

For Vibe Builders: You now have the tools to manage larger, more complex projects without getting lost in the details. Use the new team marketplace features to share your custom agent setups with others, and rely on the improved PR (pull request) review interfaces to keep your work organized as you ship faster.

For Developers: These updates allow you to exert finer control over your environment and token spend. Focus on leveraging the new subagent and MCP (Model Context Protocol) visibility to audit your agentic workflows, and use the new CLI (command-line interface) features to integrate modal editing directly into your existing terminal-based habits.

What to watch next

Watch for how these platforms handle the inevitable collision between agent autonomy and enterprise security policies. Keep an eye on whether these tools begin to standardize their plugin formats to allow for better cross-platform compatibility.