Key Signals

• GitHub Copilot SDK reaches General Availability with six language SDKs, establishing Copilot as an embeddable platform. The v1.0.0 release delivers stable APIs for Node.js, Python, Go, .NET, Rust, and Java, giving third-party developers programmatic access to the same agentic engine behind Copilot—planning, tool invocation, file edits, and multi-turn sessions. This marks a strategic shift from Copilot as a product to Copilot as infrastructure other tools can build on, directly competing with Anthropic's SDK and OpenAI's Codex SDK. [1][2]

• Copilot CLI receives its largest feature update: rubber duck critic agent, scheduled prompts, voice input, and an experimental tabbed UI. The rubber duck agent provides autonomous second-opinion code review during tasks, prompt scheduling enables recurring automated workflows (/every 30m run tests), and local speech-to-text voice input allows hands-free operation. A new experimental terminal interface with tabs for issues, PRs, and gists signals GitHub's ambition to make the CLI a complete development surface. [3][4][5]

• Cloud and local sandboxes give Copilot secure isolation infrastructure for agentic tool execution, now in public preview. Local sandboxes restrict filesystem/network access on macOS, Linux, and Windows with enterprise MDM policy enforcement. Cloud sandboxes launch ephemeral Linux environments hosted by GitHub with cross-device session continuity. This addresses the fundamental trust gap that has slowed enterprise adoption of agentic coding. [6]

• GitHub is racing to scale 30-40x as agentic coding pushes traffic far beyond 100% YoY growth projections. The platform now processes 1.4 billion commits per month (vs. 1 billion in all of 2025) and agents alone create over 17 million pull requests monthly. GitHub is migrating to Azure, rewriting core systems from Ruby to Go, and receiving Microsoft engineering reinforcements—acknowledging that traditional horizontal scaling is insufficient for AI-driven demand. [7]

• GitHub Copilot's token-based billing system goes live, replacing the premium request model with per-token pricing tied to model choice. New plans include Pro ($10/mo, $15 credits), Pro+ ($39/mo, $70 credits), and Max ($100/mo, $200 credits). Token costs vary dramatically—GPT-5.5 costs 20x more than GPT-5 mini per million input tokens—making model selection a direct cost lever. Enterprise customers get pooled credits with a new four-tier budget control system. [8]

• OpenAI's Codex discovers a critical HTTP/2 Bomb vulnerability in nginx, Apache, IIS, Envoy, and Pingora by autonomously combining two decade-old attack techniques no human had chained together. The exploit uses HPACK indexed reference bombing paired with HTTP/2 window stalling to consume and hold 32GB of server memory in 20 seconds from a single client. This is a concrete demonstration of AI agents finding novel security vulnerabilities through combinatorial reasoning across codebases. [9]

• OpenAI expands Codex beyond coding into a knowledge work platform, with 20% of its 5 million weekly active users now non-developers. New features include Sites, Annotations, and vertical plugins for data analytics, sales, product design, and investment banking—with partners including Wix, Figma, Replit, Snowflake, and Salesforce. This positions Codex in direct competition with Anthropic's Cowork. [10][11]

AI Coding News

• Microsoft announces Intelligent Terminal for Windows 11, bringing AI agents directly into the shell via the Agent Communication Protocol. The experimental release starts with GitHub Copilot integration but supports any ACP-compatible agent—Claude Code, Codex, and others. It detects errors in real-time, offers agent-driven fixes inline, and provides configurable layouts. This challenges Warp's position in the "AI-native terminal" space while leveraging Microsoft's OS-level integration advantage. [12]

• Microsoft launches Scout at Build 2026—an always-on agentic personal assistant built on the OpenClaw framework, requiring a GitHub Copilot subscription. Scout operates across Microsoft 365 with persistent memory and evolving skills, adapting to individual work patterns. It includes a policy conformance system with continuous guardrail checks and audit trails, addressing the safety concerns that emerged when OpenClaw agents famously ran amok in early 2026. Available through Microsoft's Frontier early-access program. [13]

• Microsoft open-sources Agent Control Specification, a portable policy standard for governing AI agent behavior across frameworks. ACS defines interception points where policies can allow, block, redact, or require human approval. It ships with SDK plug-ins for LangChain, OpenAI Agents SDK, Anthropic Agents SDK, AutoGen, CrewAI, Semantic Kernel, and MCP tools—making it the first cross-ecosystem agent governance standard. [14]

• Microsoft open-sources ASSERT, a framework for generating AI behavior tests from natural-language policy descriptions. ASSERT takes plain-language specs of expected behavior, generates structured test scenarios, runs them against the target system, scores results, and records intermediate tool calls for failure inspection. It addresses a gap between generic model evaluations and application-specific behavioral compliance. [15]

• Microsoft announces Rayfin, an open-source SDK/CLI for deploying vibe-coded applications into enterprise production on Microsoft Fabric, with Replit as exclusive launch partner. Rayfin generates enterprise-grade backends from code definitions and deploys them to Fabric with built-in governance. This bridges the gap between rapid AI-assisted prototyping and production-grade security/compliance that has blocked enterprise adoption of vibe coding outputs. [16]

• GitHub faces an existential reliability challenge, logging hundreds of incidents over 12 months while AI-driven traffic explodes. COO Kyle Daigle reveals the company assumed 200% growth would be the surprise scenario—actual demand requires 30-40x capacity within a year. GitHub is rewriting GitHub Actions' job dispatch system, moving performance-critical code from Ruby to Go, migrating from its own data centers to Azure, and receiving dedicated Microsoft engineering support. [7]

Feature Update

• GitHub Copilot SDK v1.0.0 reaches General Availability across Node.js, Python, Go, .NET, Rust, and Java. Key capabilities include custom tools and MCP server connections, fine-grained system prompt customization, OpenTelemetry tracing, flexible authentication, cloud/remote sessions, and a hook system for intercepting agent behavior. New since preview: Rust SDK bundles the CLI binary, multi-client workflow support, slash commands across all SDKs, and improved connection diagnostics. Available to all Copilot subscribers and non-subscribers via BYOK. [1][2]

• GitHub Copilot CLI v1.0.58 and v1.0.59 ship with rubber duck enabled by default, Remote JSON RPC enabled, and the /voice command for local speech-to-text dictation. The v1.0.58 release also adds experimental features behind /experimental on: scheduled prompts via /every and /after, a new GitHub color theme, and the redesigned UI with tabs for issues, PRs, and gists. The v1.0.59 release adds the /voice slash command. [3][4][5]

• Cloud and local sandboxes for GitHub Copilot enter public preview, providing isolated execution environments for agentic tool use. Local sandboxes use Microsoft MXC technology for cross-platform filesystem/network restriction with Intune policy enforcement. Cloud sandboxes offer ephemeral GitHub-hosted Linux environments inheriting existing cloud agent policies. Cross-device session continuity and parallel compute offloading are supported in cloud mode. [6]

• Copilot code review gains agent skills, MCP server connections, and a new Medium analysis tier in public preview. Custom skills invoke team-specific tools during reviews; MCP pulls context from issue trackers, documentation, and service catalogs. The Medium tier routes complex PRs to a higher-reasoning model for deeper analysis of security-sensitive and cross-service changes, while Low remains fast and cost-efficient. Configuration is shared with cloud agent for single-investment platform setup. [17]

• GitHub launches Agent Apps—partner AI agents installable from the Marketplace and integrated directly into GitHub workflows. Agents can be assigned issues, @mentioned in PR comments, or prompted via the Agents UI. Launch partners include Amplitude, Bright Security, Endor Labs, LaunchDarkly, Miro, Sonar, PagerDuty, Packfiles, and Octopus Deploy. A waitlist is open for developers wanting to build their own agent apps. [18]

• The GitHub Copilot app technical preview expands to all existing Copilot Pro, Pro+, Business, and Enterprise customers on Windows, macOS, and Linux. This standalone application provides a dedicated Copilot experience outside of IDE integrations and the CLI. [19]

• GPT-4.1 is deprecated across all GitHub Copilot experiences including Chat, inline edits, ask and agent modes, and code completions. The deprecation took effect June 1, 2026. Users are directed to suggested alternatives for each use case. [20]

• Gemini 3.1 Pro and Gemini 3.5 Flash are now available in Copilot CLI, Copilot cloud agent, and the GitHub Copilot app. This expands Google's model presence across additional Copilot surfaces beyond the initial integration. [21]

• Copilot CLI and agentic capabilities launch in GitHub Copilot for JetBrains IDEs, bringing CLI sessions and broader agentic improvements to the JetBrains ecosystem. This update centers on new capabilities for Copilot CLI sessions within the IDE while also delivering agentic enhancements. [22]

• GitHub Copilot code review for Azure Repos enters technical preview, bringing on-demand PR reviews directly into Azure DevOps workflows. This extends Copilot code review beyond GitHub-hosted repositories to Azure DevOps customers. [23]

• GitHub Copilot's usage-based billing goes live with token-based pricing and new plan tiers. Pro ($10/mo) receives $15 credits, Pro+ ($39/mo) gets $70, and the new Max tier ($100/mo) provides $200 credits for heavy agentic use. Business ($19/seat) and Enterprise ($39/seat) use pooled organization credits. Model costs range from $0.25/M tokens (GPT-5 mini) to $5/M tokens (GPT-5.5, Opus 4.8). Code completions and next edit suggestions remain free. A four-tier budget control system governs overage. [8]

• Claude Code v2.1.161 improves parallel tool execution and observability. Failed Bash commands no longer cancel other calls in the same batch—each tool returns independently. OTEL_RESOURCE_ATTRIBUTES values are now metric labels for slicing by team/repo. Additional fixes address managed-settings policies blocking third-party providers, background subagent output corruption, worktree isolation, and stale model selection in background sessions. Terminal rendering performance improved via layout engine JIT stabilization. [24]

• Claude Code v2.1.160 adds security prompts for shell startup files and build-tool configs in acceptEdits mode. Writing to .zshenv, .bash_login, ~/.config/git/, .npmrc, .bazelrc, .pre-commit-config.yaml, and .devcontainer/ now requires explicit approval. The read-before-edit check is relaxed: single-file grep commands now satisfy it. Multiple fixes address WSL clipboard, session restoration, background daemon reliability, Windows responsiveness under load, and CJK IME positioning. The dynamic-workflow trigger keyword is renamed from "workflow" to "ultracode." [25]

• OpenAI Codex launches Sites, Annotations, and vertical plugins for knowledge workers. Sites create shareable interactive dashboards and apps via workspace URLs, with partners Wix, Base44, Replit, Lovable, Figma, and Emergent. Annotations let users target specific document sections for AI edits. New plugins cover data analytics, sales, product design, and investment banking. [10][11]

• Gemini CLI ships nightly v0.45.0-nightly.20260602 with a transition to the Flash GA model when an experiment flag is present. This is a minor nightly release focused on model routing updates. [26]