Key Signals

• GitHub launches the Copilot App as GA alongside Agent Finder and Auto Mode, marking its most concentrated push into agent-native development. The standalone desktop app provides parallel agent sessions via git worktrees, bidirectional canvases, cloud automations, and sandboxed execution — positioning GitHub as a mission control for coding agents rather than just a code completion tool. Agent Finder implements the new open Agentic Resource Discovery specification so Copilot can dynamically discover and rank MCP servers, skills, and tools without preloading them into context. Auto Mode in Copilot Chat adds intelligent model routing across Claude Sonnet 4.6, GPT-5.4, and Haiku 4.5 with a 10% token discount. [1][2][3]

• Cursor ships cloud subagents that run in isolated VMs, enabling parallel agentic workloads without tying up local machines. The new /in-cloud command spins up a cloud subagent on its own branch and VM, and the /babysit command lets a cloud agent iterate on a PR until it's merge-ready. Sessions can be handed off between local and cloud seamlessly, and reusable environment snapshots speed up future agent starts for the entire team. [4]

• AWS launches Kiro Mobile for iOS, letting developers start, steer, and approve autonomous coding sessions from their phones. The native iOS app supports Chat, Spec, and Autonomous session modes, with cloud compute continuing sessions even when the screen goes dark. AWS frames spec-driven development — where agents produce requirements and design documents before writing code — as the antidote to "AI coding slop," and reports 80% of its software engineers now use Kiro. [5]

• Block reveals BuilderBot merges 1,500 PRs per week (~15% of all production code) via AI agents orchestrated from Slack. Built on the open-source Goose framework, BuilderBot picks up tickets from Linear and Jira, handles cross-service changes, watches CI, and iterates on feedback — all within Slack threads. The disclosure shows the scale at which AI-native engineering is already operating at major fintech companies, with engineers managing agents rather than writing every line. [6]

• AWS DevOps Agent adds release readiness review and autonomous release testing, targeting the delivery bottleneck created by AI-generated code. Rather than another coding assistant, these features evaluate cross-repository dependency risks, run the software in isolated environments, and generate change-specific test plans — acknowledging that writing code is no longer the constraint; getting it safely to production is. [7]

• The Trump administration's unprecedented use of export controls against Anthropic's Fable 5 and Mythos 5 creates regulatory uncertainty for all frontier AI model providers. Experts call it the first time US export controls have targeted access to a cloud-hosted AI model, and note the order does not fit neatly into existing frameworks since no model weights or code are actually being transferred. The episode raises questions about what protections will be considered sufficient for the next generation of models from any frontier lab. [8]

AI Coding News

• Vercel launched "eve," an open-source Apache 2.0 framework described as "Next.js for agents" at its Ship conference in London. Eve defines agents as directories of files — model config, Markdown system prompts, and TypeScript tool files where filenames become tool names — then compiles them into durable workflows with crash recovery via Vercel's Workflow SDK. Each agent gets sandboxed execution, subagent delegation, MCP server support, and OpenTelemetry tracing. Vercel reports running 100+ agents internally and says agents now trigger 29% of its platform deployments, up from less than 3% a year ago. [9]

• Google, Microsoft, and OpenAI are among 13 founding members of the Appia Foundation, a Linux Foundation initiative to build AI's missing trust and compliance verification layer. Appia will produce open "conformity specifications" that translate international standards into concrete, assessable criteria with "evidence pass-through" for multi-party AI supply chains. The foundation addresses a gap where AI trustworthiness today is largely self-declared, with no standardized verification mechanism. [10]

• NVIDIA's GEAR lab demonstrated ENPIRE, an agent harness framework that lets AI coding agents autonomously direct overnight robot training without human intervention. Tested with OpenAI Codex (GPT-5.5), Claude Code (Opus 4.7), and Kimi Code (K2.6), the coding agents independently developed training approaches that taught robots to insert GPUs into motherboard sockets and cut zip ties. Jim Fan, NVIDIA's director of AI, said the team plans to open-source everything so anyone can host their own "self-running robot lab at home." [11]

Feature Update

• GitHub Copilot App reaches general availability for macOS, Windows, and Linux as a standalone desktop application for agent-driven development. Each agent session runs in its own git worktree for isolation, with new canvases providing bidirectional work surfaces for plans, terminals, and browser sessions that both humans and agents can update. Cloud automations allow scheduling recurring agent work without local machine dependency, and sessions support bring-your-own-model via MCP server connections. The Copilot SDK is now GA, exposing the same engine for custom agentic tooling. [1][12]

• GitHub Agent Finder enables dynamic discovery of MCP servers, skills, agents, and tools via the open Agentic Resource Discovery specification. Instead of preloading every available tool into context, Copilot now searches a registry at task time and returns ranked matches. Enterprises can scope discovery to private registries and enforce governance through managed settings. No auto-installation occurs — the developer approves what gets wired in. Available on all Copilot plans. [2]

• Auto Mode in GitHub Copilot Chat is now generally available on github.com and the GitHub mobile app for all plans. Auto dynamically routes to models like Claude Sonnet 4.6, GPT-5.4 mini, GPT-5.4, and Haiku 4.5 based on request complexity and real-time availability, with transparent model attribution on each response. Usage is billed at a 10% discount for paid subscribers. [3]

• Enterprise-managed settings now support disableBypassPermissionsMode to prevent Copilot CLI and VS Code from auto-approving commands. This is the first governance capability in GitHub's enterprise-managed settings configuration, allowing admins to block "yolo mode" across their organization via .github-private/.github/copilot/settings.json. VS Code v1.122+ already respects this policy. [13]

• GitHub is reopening sign-ups for Copilot Student, Pro, Pro+, and Max plans with new upgrade and pay-as-you-go options. Users approaching limits can upgrade mid-cycle or pay for additional consumption without switching plans. Roll-out is gradual over two weeks. [14]

• Copilot CLI v1.0.64-0 adds the /diagnose command, MCP registry browsing, and makes /security-review generally available. The release also adds plugin-provided MCP server discovery, CSV output for MCP tools, and a /loop alias. Bug fixes address image persistence across sessions, Azure CLI credential handling, Windows glob patterns, and content exclusion filtering. The legacy intent-reporting tool is removed. [15]

• Claude Code v2.1.181 introduces /config key=value syntax for setting any configuration from the prompt, upgrades bundled Bun to 1.4, and adds auto-retry for API connection drops. The release fixes prompt caching failures on custom ANTHROPIC_BASE_URL and Foundry deployments, resolves Write/Edit producing 0-byte files on network drives and cloud-synced folders, and fixes a ~120ms startup regression introduced in v2.1.169. Subagent UX improves with idle auto-hide, capped list rows, and proper depth limiting. [16]

• Cursor introduces cloud environment setup and cloud subagents in the Agents Window. Dev environments can be set up in under 10 minutes with reusable snapshots captured in .cursor/environment.json. The /in-cloud command runs subagents on isolated VMs and branches, while /babysit iterates remotely to prepare PRs for merge. Sessions can be moved between local and cloud reliably. [4]

• Kiro launches a native iOS app for mobile supervision of agentic coding sessions, announced at AWS Summit New York. The app supports the same three session modes as Kiro Web, with native diff rendering and PR status on every session row. Available in preview on iOS 26+ for Kiro Pro and above subscribers; Android is not currently planned. [5]

• AWS DevOps Agent gains release readiness review and autonomous release testing in preview. The readiness review checks cross-repository dependency risks and compliance against plain-English standards, outputting BLOCK/Caution/Safe verdicts as PR comments in GitHub or GitLab. Autonomous release testing generates change-specific test plans rather than running static suites. Available at no additional cost in US East. [7]

• Gemini CLI v0.48.0-preview.0 ships with standardized tool output formatting, zero-quota fail-fast behavior, and a security fix preventing path traversal during skill installation. Other fixes address multi-line escaped quotes in shell wrappers, tmux false-positive background detection, pending tools and trust overrides, and GDC air-gapped Service Identity support. The coreTools config setting migrates to tools.core. [17]

• OpenCode v1.17.8 delivers faster session timeline loading and multiple MCP compatibility fixes. OpenAI-compatible providers now accept previously failing MCP tool schemas, Cloudflare AI Gateway receives API keys correctly, long-running MCP tools maintain timeouts on progress, and MCP tool failures surface server error text instead of generic messages. Desktop adds a Home tab toggle and a faster v2 file picker. [18]

• Copilot SDK publishes go/v1.0.2 with refreshed Azure managed identity documentation for Microsoft Foundry. The update converts language examples to tabs and adds environment configuration and branding guidance for Azure Identity. [19]

• OpenAI Codex publishes three alpha pre-releases (0.141.0-alpha.4 through alpha.6) on June 17 as part of active development toward the 0.141.0 milestone. The full 0.141.0 release (arriving June 18) brings end-to-end encrypted Noise relay channels for remote executors, cross-platform remote execution preserving native paths, plugin-activated MCP servers per thread, and TUI auto-resolution timers. [20]