Key Signals

• Copilot CLI breaks free from the terminal with remote session control. GitHub launched copilot --remote in public preview, letting developers monitor and steer running CLI sessions from the GitHub web UI or mobile apps in real time. Combined with the v1.0.25 release that adds MCP server registry installation with guided configuration and ACP client MCP provisioning, this marks a fundamental shift from Copilot CLI as a local-only tool to a remotely accessible agentic platform. Enterprise teams can now delegate long-running agent tasks and check in from any device. [1][2]

• Claude Mythos Preview scores 93.9% on SWE-bench Verified, but Anthropic withholds public access. Anthropic's most capable frontier model represents a 13-point leap over Claude Opus 4.6 (80.8%) on the key coding benchmark, and during internal testing it autonomously discovered zero-day vulnerabilities in every major OS and browser. Rather than a general release, access is restricted to a ten-company consortium — including AWS, Apple, Google, Microsoft, and NVIDIA — through Project Glasswing, backed by $100M in Anthropic usage credits. This unprecedented restricted-release strategy signals that frontier coding capability is increasingly viewed as a dual-use security concern. [3]

• Kiro CLI 2.0 ships Windows support, headless CI/CD mode, and graduates its terminal UI. Amazon's Kiro CLI now runs natively on Windows 11, joining macOS and Linux with full feature parity including custom agents, MCP servers, and the new default terminal UI. Headless mode enables non-interactive execution in CI/CD pipelines — a direct play for the same automated-agent-in-pipeline use case that Copilot CLI and Claude Code are also pursuing. Subagents now support task dependency chains with parallel execution of independent steps. [4]

• GitHub Copilot gains data residency for US and EU with FedRAMP compliance. All generally available Copilot features — agent mode, chat, cloud agent, code review, PR summaries, and CLI — now support region-locked inference with FedRAMP Moderate authorization for US government customers. Models at launch include GPT-5.4, Claude Sonnet 4.6, and Claude Opus 4.6, with a 10% model multiplier surcharge for data-resident requests. This removes a major blocker for regulated industries and government agencies evaluating AI coding tools. [5]

• Microsoft is building an OpenClaw-like always-on agent for Microsoft 365 Copilot. The enterprise-focused agent would bring OpenClaw's multistep, long-running task capabilities into the 365 ecosystem with stricter security controls. This adds to Microsoft's growing agentic portfolio that already includes Copilot Cowork and Copilot Tasks, and is expected to be showcased at Build in June. The move underscores how the open-source OpenClaw project is reshaping enterprise expectations for what AI agents should do. [6]

• Cursor 3.1 introduces tiled agent layout for parallel multi-agent workflows. The new tiled pane system lets developers run and compare multiple agents side-by-side with persistent layout across sessions, alongside upgraded batch-STT voice input and branch selection for cloud agents before launch. These UX improvements reflect the industry's rapid evolution from single-agent chat to orchestrating multiple concurrent agentic sessions as a standard developer workflow. [7]

AI Coding News

• Anthropic restricts Claude Mythos Preview to a ten-company consortium via Project Glasswing. Claude Mythos Preview achieved a 93.9% score on SWE-bench Verified, a dramatic improvement over Opus 4.6's 80.8%, and autonomously discovered zero-day vulnerabilities across every major OS and browser during internal testing — including a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg's H.264 codec. On the OSS-Fuzz corpus, it achieved full control flow hijack on ten separate, fully patched targets. Rather than public release, Anthropic launched Project Glasswing with AWS, Apple, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, committing $100M in usage credits to identify and patch vulnerabilities in critical software. The system card and risk report are publicly available, and findings will inform future Claude releases. [3]

• Microsoft confirms development of an OpenClaw-inspired agent for 365 Copilot enterprise customers. The agent would function as an always-on version of Microsoft 365 Copilot capable of completing multistep tasks over long periods, addressing the same autonomous workflow niche that has made OpenClaw popular among developers. Microsoft already offers Copilot Cowork and Copilot Tasks, both of which run in the cloud; the new agent may adopt local execution features that OpenClaw advocates favor. The Mac Mini has been the go-to hardware for OpenClaw users, giving Microsoft additional motivation to offer a native Windows alternative. A reveal is expected at Microsoft Build in June. [6]

• Cloudflare Agent Cloud integrates OpenAI's GPT-5.4 and Codex for enterprise agentic workflows. The partnership brings OpenAI's models to Cloudflare's edge infrastructure, enabling enterprises to build, deploy, and scale AI agents for real-world tasks with speed and security. This positions Codex not only as a standalone CLI coding tool but also as an embeddable engine for enterprise agent platforms. [8]

Feature Update

• GitHub Copilot CLI v1.0.25 adds MCP server registry installation, remote sessions, and new diagnostics. Users can now install MCP servers from the registry with guided configuration directly in the CLI, and ACP clients can provide MCP servers when starting or loading sessions. The new /env command displays loaded environment details including instructions, MCP servers, skills, agents, and plugins. Remote CLI session support via --remote or /remote allows controlling sessions from GitHub web and mobile. Additional fixes address session history persistence, skill instruction retention across turns, and MCP remote connection auto-retry on transient failures. [2]

• Copilot CLI remote session control enters public preview on web and mobile. copilot --remote streams CLI session activity to GitHub in real time, displaying a link and QR code for access from another device. Remote sessions support full feature parity: mid-session steering, plan review and modification, mode switching, permission approval, and ask_user responses. Each session is private to the user who started it, and enterprise administrators must explicitly enable remote control policies. [1]

• GitHub Copilot data residency launches for US and EU with FedRAMP Moderate authorization. All GA features are supported with region-locked inference, including agent mode, inline suggestions, chat, cloud agent, code review, PR summaries, and CLI. Launch models include GPT-5.4, Claude Sonnet 4.6, and Claude Opus 4.6 from both OpenAI and Anthropic; Gemini models are excluded pending GCP data-resident endpoints. Data-resident and FedRAMP requests carry a 10% model multiplier surcharge, and additional regions are planned for later in 2026. [5]

• Copilot cloud agent can now fix merge conflicts in three clicks on github.com. A new "Fix with Copilot" button appears on PRs with merge conflicts; clicking it prepopulates a comment asking Copilot to resolve conflicts, verify builds and tests, and push — all from a cloud-based development environment. Users can also @copilot in PRs to fix failing Actions workflows, address code review comments, or make arbitrary changes. Available with all paid Copilot plans. [9]

• Claude Code v2.1.105 adds PreCompact hooks, plugin monitors, and 20+ bug fixes. New features include the path parameter for EnterWorktree, PreCompact hook support that can block compaction (exit code 2 or {"decision":"block"}), and background monitor support for plugins via a monitors manifest key. The skill description listing cap was raised from 250 to 1,536 characters, WebFetch now strips <style> and <script> tags to preserve content budgets, and stalled API streams auto-abort after 5 minutes of no data with a non-streaming retry. Over 20 fixes address MCP tool availability on headless sessions, Bedrock model picker region issues, 429 rate-limit error display, marketplace auto-update corruption, and terminal color palette rendering across Ghostty, Kitty, Alacritty, and other modern terminals. [10]

• Cursor 3.1 ships tiled agent layout, upgraded voice input, and branch selection for cloud agents. The tiled pane system allows splitting the view to run multiple agents in parallel with drag-and-drop tile management and persistent layouts across sessions. Voice input now uses batch STT for more reliable transcription, and a new branch selector in the empty state lets users target cloud agents at a specific branch before launch. Additional improvements include diff-to-file navigation and include/exclude filters for file search. [7]

• Kiro CLI 2.0 launches with Windows 11 support, headless mode, and terminal UI as default. The terminal UI provides syntax-highlighted markdown, interactive overlay panels, visual tool progress, and a crew monitor for tracking subagent activity. Headless mode authenticates via KIRO_API_KEY and executes prompts end-to-end with --no-interactive, supporting --trust-all-tools for permission preauthorization. Subagents now support task dependency chains — e.g., analyze a codebase, refactor modules, then run tests — with parallel execution of independent steps. Available for Pro, Pro+, and Power subscribers, with enterprise governance controls for API key generation. [4]

• OpenAI Codex publishes three v0.121.0 alpha pre-releases building toward the next stable version. Versions alpha.4, alpha.5, and alpha.6 were released in sequence on April 13, following the v0.120.0 stable release on April 11 that introduced Realtime V2 background agent streaming, improved hook activity TUI rendering, Windows elevated sandbox fixes, and Guardian review improvements. No individual changelogs accompany the alpha builds. [11]

• Gemini CLI v0.37.2 ships a patch release with a targeted bug fix. The release cherry-picks commit 9d741ab onto v0.37.1, addressing a specific issue identified in PR #24565. This is a minor maintenance release with no feature changes. [12]