Key Signals

• Anthropic's Claude Sonnet 4.6 launches with near-Opus coding performance at a fraction of the price, triggering immediate adoption across the AI coding tool ecosystem. The new model scores within a percentage point of Opus 4.6 on coding benchmarks while priced at $3/$15 per million tokens versus Opus's $5/$25. Developers in early access preferred Sonnet 4.6 over the older Opus 4.5 model 59% of the time, with notably fewer hallucinations and less over-engineering. Within hours, GitHub Copilot, Copilot CLI, Claude Code, and Kiro all shipped Sonnet 4.6 support, signaling a new baseline for mid-tier agentic models. [1][2][3]

• Cursor 2.5 launches a Plugin Marketplace and async subagents, establishing a platform ecosystem model for AI coding editors. The marketplace packages skills, subagents, MCP servers, hooks, and rules into single installable plugins, with launch partners including Amplitude, AWS, Figma, Linear, and Stripe. Async subagents represent a significant architectural shift — parent agents can now continue working while spawning nested trees of background subagents, enabling much larger multi-file refactors and complex bug fixes. The addition of granular sandbox network controls also addresses enterprise security requirements for agent-initiated network access. [4]

• GitHub expands Copilot coding agent access surfaces to Visual Studio and Raycast, while Eclipse gets MCP Registry and agent task management. The autonomous Copilot coding agent — which creates draft PRs and works in the cloud — is now launchable from Visual Studio 2026 via a dedicated button in Copilot Chat, and from Raycast using the new GitHub Copilot extension. Meanwhile, Copilot in Eclipse gains an MCP Registry for one-click MCP server discovery and installation, plus a real-time Todo List tool that tracks agent progress on complex tasks. These expansions push agentic coding workflows beyond VS Code into the broader IDE landscape. [5][6][7]

• Copilot CLI v0.0.411 makes Autopilot mode and the /fleet command generally available to all users. This release also adds SDK APIs for plan mode, autopilot, fleet, and workspace files, significantly expanding the programmatic surface for building on top of Copilot CLI. New support for MCP servers from the Windows On-Device Registry and the include_coauthor config option for controlling Co-authored-by git trailers round out a substantial update that improves both the autonomous and integrations story. [9]

• OpenAI publishes the Codex App Server architecture, revealing that they tried and rejected MCP for IDE integration. The bidirectional JSON-RPC protocol now powers every Codex surface — CLI, VS Code, web, macOS desktop, Xcode, and JetBrains — through Item/Turn/Thread conversation primitives. OpenAI found that MCP's tool-oriented model couldn't accommodate streaming diffs, approval flows, and thread persistence needed for rich IDE experiences, opting instead for a purpose-built protocol. Codex is also listed among ACP-compatible agents, the emerging standard initiated by Zed Industries. [10]

• Gemini CLI v0.30.0-preview.0 ships a formalized 5-phase planning workflow, SDK bootstrap, and defaults to Gemini 3. This massive preview release with 100+ commits introduces Plan Mode with full MCP server support, a new --policy flag replacing the deprecated --allowed-tools, and tool output masking enabled by default for security. The initial SDK package includes SessionContext for tool calls, dynamic system instructions, and custom skills — positioning Gemini CLI as a programmable platform rather than just a terminal tool. [11]

• Moonshot AI's Kimi K2.5 introduces Agent Swarm mode with up to 100 parallel sub-agents, trained via a novel RL technique called PARL. The open-weight multimodal model achieves coding benchmarks comparable to GPT-5 and Gemini, while its Agent Swarm outperformed GPT-5.2 Pro on BrowseComp and Claude Opus 4.5 on WideSearch. The Parallel Agent Reinforcement Learning approach freezes subagents and trains only the orchestrator, addressing serial collapse and credit assignment challenges. This represents a significant step toward production-ready multi-agent orchestration in open-weight models. [12]

AI Coding News

• Microsoft AI CEO Mustafa Suleyman predicts all white-collar desk work will be automated within 18 months. Suleyman expects "human-level performance on most, if not all professional tasks" from AI, specifically targeting work involving "sitting down at a computer" — accounting, legal, marketing, and project management. He pointed to exponential growth in computational power and predicted that creating a new AI model will soon be as easy as "creating a podcast or writing a blog." While the timeline is aggressive, the statement signals Microsoft's strategic framing of AI agents as replacements for, not just augments to, knowledge work. [13]

• Anthropic releases Claude Sonnet 4.6 with a 1M token context window and record ARC-AGI-2 score of 60.4%. The new model doubles the context window previously available for Sonnet, which Anthropic describes as "enough to hold entire codebases, lengthy contracts, or dozens of research papers in a single request." On office tasks measured by OpenAI's GDPval benchmark, Sonnet 4.6 even outperforms Opus 4.6, Gemini 3 Pro, and GPT-5.2. The model supports context compaction and adaptive thinking for extended thinking token budgets, and becomes the default on claude.ai for Free and Pro plans. An updated Haiku model is expected to follow in coming weeks. [1][2][3]

• InfoQ details the Codex App Server's conversation primitives and deployment patterns for unifying agent surfaces. The architecture defines Items, Turns, and Threads. Three deployment patterns serve different integration needs: local binary bundling for VS Code, decoupled release cycles for partner integrations like Xcode, and container-based runtimes for the web app. All source code is available in the open-source Codex CLI repository. [10]

Feature Update

• GitHub Copilot now offers Claude Sonnet 4.6 across all surfaces including VS Code, Visual Studio, github.com, GitHub Mobile, Copilot CLI, and Copilot Coding Agent. The model excels at agentic coding and search operations, launching with a tentative 1x premium request multiplier. Copilot Pro, Pro+, Business, and Enterprise users can access it via the model picker, though Business and Enterprise administrators must first enable the Claude Sonnet 4.6 policy in Copilot settings. Rollout is gradual. [8]

• GitHub Copilot in Eclipse gains MCP Registry, Todo List management, editor selection context, and refreshed Chat UX. The MCP Registry enables one-click discovery and installation of MCP servers without manual configuration. The Todo List tool lets the agent create and track structured task lists in real-time during complex operations. Chat view improvements include font size control via keyboard shortcuts, a refreshed dark theme, undo/redo support, and native toolbar integration. Editor Selection Context automatically includes selected code in chat context. [7]

• Copilot coding agent is now accessible from Visual Studio 2026 via a dedicated "Send to Copilot Coding Agent" button in Copilot Chat. The asynchronous background agent opens a draft pull request, works in the cloud, and requests a review when done. Users need Visual Studio 2026 December Update 18.1.0 or later and must enable the "Enable Copilot Coding agent" setting. Available to all Copilot subscription tiers, with Business and Enterprise requiring administrator enablement. [5]

• Copilot coding agent can now be assigned issues from Raycast on Windows and macOS. The GitHub Copilot Raycast extension supports "Assign Issue to Copilot," "Create Task," and "View Tasks" commands. Users select a repository and issue, press Ctrl+Enter or Cmd+Enter, and Copilot opens a PR and works autonomously in the background. This brings agentic coding workflows into the productivity launcher ecosystem. [6]

• Copilot CLI v0.0.411 ships Autopilot mode GA, Claude Sonnet 4.6 support, and SDK APIs for plan/fleet/workspace. The /fleet command is now available to all users, and new SDK APIs expose plan mode, autopilot, fleet, and workspace file management for programmatic integration. Additional highlights include MCP server support from the Windows On-Device Registry, the include_coauthor config option to control Co-authored-by trailers in git commits, improved memory storage display with subject/fact/citations in the timeline, and numerous alt-screen mode improvements including reduced memory usage during long sessions. [9]

• Cursor 2.5 launches Plugin Marketplace, sandbox network access controls, and async subagents. Plugins bundle skills, subagents, MCP servers, hooks, and rules into single installs, with initial partners including Amplitude, AWS, Figma, Linear, and Stripe. Sandbox network access controls let users define exactly which domains the agent can reach, with Enterprise admins able to enforce organization-wide egress policies. Async subagents can now spawn nested subagent trees for parallel execution of large tasks, with improved latency, streaming feedback, and responsiveness. [4]

• Claude Code v2.1.45 adds Sonnet 4.6 support, SDK rate limit events, and fixes critical Agent Teams and sandbox issues. New SDKRateLimitInfo and SDKRateLimitEvent types enable SDK consumers to receive rate limit status updates including utilization and reset times. Key fixes resolve Agent Teams failures on Bedrock, Vertex, and Foundry, macOS sandbox "operation not permitted" errors for temp files, and backgrounded agent crashes on completion. Startup performance improves via deferred session history loading, and memory usage for shell commands with large output no longer grows unboundedly. [14]

• Codex v0.102.0 introduces unified permissions flow, customizable multi-agent roles, and structured network approvals. The permissions overhaul provides clearer history in the TUI and a slash command for granting sandbox read access to blocked directories. Structured network approval prompts now show richer host/protocol context directly in the approval UI. Customizable multi-agent roles via config represent a step toward Codex's multi-agent architecture, and a new model/rerouted notification lets clients detect and render model reroute events. Multiple js_repl stability fixes address reset hangs, in-flight tool-call races, and view_image panics. [15]

• Codex v0.103.0 adds richer app listing details and commit co-author attribution with configurable overrides. App listing responses now include app_metadata, branding, and labels for more complete app card rendering without extra requests. Commit co-author attribution uses a Codex-managed prepare-commit-msg hook, with command_attribution override support allowing users to set a default label, custom label, or disable attribution entirely. [16]

• Gemini CLI v0.30.0-preview.0 delivers Plan Mode, SDK foundation, and Gemini 3 as default model in a 100+ commit release. Plan Mode introduces a formalized 5-phase sequential planning workflow with MCP server support, plan editing via the replace tool, and session-isolated plan files. The SDK foundation includes SessionContext for tool calls, dynamic system instructions, and custom skills. A new --policy flag replaces the deprecated --allowed-tools and excludeTools, tool output masking is enabled by default, and sub-agents transition to XML format. UX additions include inline thinking bubbles, DevTools integration via F12, Vim motions, and Solarized themes. [11]

• Kiro v0.9.47 adds Claude Sonnet 4.6 support and fixes enterprise account handling for extensions. Kiro joins GitHub Copilot, Copilot CLI, and Claude Code in same-day adoption of Anthropic's latest mid-tier model, reflecting how quickly the ecosystem responds to new model releases. The enterprise account fix ensures extensions work correctly with organizational authentication. [17]