Key Signals

• GitHub launched the Copilot app in technical preview, its first native desktop experience purpose-built for agentic development. The app lets developers start sessions directly from issues, PRs, and previous sessions, work across isolated branches, validate with an integrated terminal and browser, and ship through standard PR review. Agent Merge can autonomously address review comments, fix failing checks, and merge once conditions are met — positioning GitHub's agentic workflow as an end-to-end loop rather than just a code generator. [1]

• OpenAI brought Codex into the ChatGPT mobile app, enabling full remote management of coding agents from iOS and Android. Unlike Anthropic's Dispatch, Codex mobile provides cross-thread access to all active environments, letting developers review outputs, approve commands, change models, and start new tasks from their phone. With 4+ million weekly users reported, this also introduced remote SSH connections to enterprise environments and HIPAA-compliant usage for Enterprise customers. [5][6][7]

• Anthropic announced a new monthly Agent SDK credit pool that separates programmatic from interactive Claude usage, effective June 15. Pro users receive $20/month, Max 5x gets $100, and Max 20x gets $200. This reverses last month's policy that blocked subscription coverage for third-party tools — now tools like Conductor and OpenClaw draw from the dedicated credit. The move signals Anthropic trying to balance ecosystem openness with sustainable unit economics as programmatic usage scales. [8]

• Claude Code v2.1.142 upgraded Fast mode to Opus 4.7 by default and added extensive agent configuration flags. New --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, and --dangerously-skip-permissions flags enable full programmatic configuration of dispatched background sessions. Critical fixes address daemon stability after macOS sleep/wake and background agents crash-looping — issues that blocked reliable unattended operation. [13]

• The AI coding agent market is rapidly shifting toward persistent cloud execution, with Conductor's $22M-backed Cloud launch as the latest signal. Conductor Cloud runs coding agents in hosted environments that persist after developers disconnect, joining Anthropic's Managed Agents, Mistral's Vibe cloud, and Roo Code's Roomote in the trend away from local-only tooling. The emerging paradigm is developers supervising fleets of longer-running agents across laptops, terminals, browsers, and hosted infrastructure. [9]

• GitHub Copilot's CLI agent arrived in JetBrains IDEs with worktree isolation and a unified sessions dashboard. Developers can delegate tasks from JetBrains to a locally running Copilot CLI agent, choosing between worktree isolation or workspace isolation. The unified sessions view tracks all agent sessions with status, elapsed time, and agent-type filtering — reflecting the shift toward managing multiple concurrent agents as a primary IDE workflow. [4]

AI Coding News

• Anthropic introduced separate Agent SDK credit pools that fundamentally restructure how programmatic Claude usage is billed. Starting June 15, claude -p, Claude Code GitHub Actions, Agent SDK usage in Python/TypeScript, and third-party tool authentication all draw from a dedicated monthly credit instead of general subscription limits. Credits don't roll over, can't be pooled across teams, and overflow to pay-as-you-go API rates. Developer Platform API-key accounts remain on standard pay-as-you-go billing unchanged. [8]

• AI coding agents are migrating from local terminals to persistent cloud environments, with Conductor Cloud as the latest example of this industry-wide shift. Conductor's Series A–backed cloud offering runs Claude Code and Codex agents in hosted workspaces that continue operating after developers disconnect. This joins Anthropic's Managed Agents, Mistral's Vibe cloud agent, and Roo Code's Roomote in abandoning local-only execution. The key challenge is the "interface problem" — developers can effectively manage 3–5 concurrent agents but need better orchestration UIs to scale beyond that. [9]

• InfoQ detailed Anthropic's postmortem on six weeks of Claude Code quality complaints traced to three overlapping product-layer changes. A March 4 reasoning effort downgrade, a March 26 caching bug that progressively erased thinking sections on every turn, and a system prompt verbosity limit each degraded different slices of traffic independently. The API and model weights were never affected. All issues were resolved April 20 (v2.1.116), with Anthropic implementing staged rollouts and automated regression testing to prevent recurrence. [10]

• An open-source hardware project called Clawdmeter brings Claude Code usage stats to a tiny AMOLED desktop dashboard. The Waveshare ESP32-S3 device connects over Bluetooth, shows pixel-art Clawd animations that intensify with usage rate, and displays session/weekly utilization charts. Side buttons trigger Claude Code's voice mode and mode-toggle shortcuts. The project amassed 800+ GitHub stars within four days of launch, reflecting the "tokenmaxxing" culture where engineers maximize AI token consumption as a productivity signal. [11]

• Lovable backed Danish hardware startup Atech's $800K pre-seed round to bring "vibe coding" to physical prototyping. Users buy a starter hardware kit, describe their desired prototype to an AI chatbot, and receive generated code to build it. The user base spans from children building cars to industrial hydrogen synthesis plants. A16z's scout fund and Sequoia Scout Fund also participated, betting that hardware's accessibility gap will collapse the way software's did. [12]

• Sea Limited's CPO explained why the company is deploying Codex across engineering teams to accelerate AI-native software development in Asia. The partnership represents a major enterprise adoption signal for OpenAI's coding tools in the APAC region, where Sea operates Shopee, Garena, and SeaMoney across Southeast Asian markets. [18]

Feature Update

• GitHub Copilot app launched in technical preview as a native desktop experience for end-to-end agentic development. Sessions start from GitHub artifacts, each with isolated branches and file state. The integrated terminal and browser enable validation without context-switching, and Agent Merge automates PR follow-through including addressing review comments and fixing CI failures. Pro/Pro+ subscribers get early access now; Business/Enterprise rollout follows this week. [1]

• Copilot cloud agent now supports auto model selection with a 10% discount on multiplier rates. Selecting "Auto" in the model picker lets Copilot intelligently choose the best available model based on system health and performance. Users opting in are exempt from weekly rate limits, creating a strong incentive to delegate model selection. [2]

• GitHub exposed team-level Copilot usage metrics via two new REST API endpoints. The user-teams-1-day report maps each Copilot-licensed user to their teams; joining this with the per-user report produces team-level aggregates covering IDE completions, chat, CLI, code review, and cloud agent activity. Breakdowns by language, IDE, feature, and model are available. Teams with fewer than five Copilot-seated users are excluded for privacy. [3]

• GitHub Copilot's CLI agent is now available in JetBrains IDEs in public preview with worktree and workspace isolation modes. The unified sessions view shows live status, elapsed time, and agent type for all running sessions. New features include an Ask question tool for clarifying ambiguity, global .agent.md support in ~/.copilot/agents, and GHES sign-in support. Edit mode has been deprecated in this release. [4]

• OpenAI Codex is now available in the ChatGPT mobile app with full cross-thread environment management. The mobile client connects to desktop/remote Codex instances through a secure relay layer, inheriting all capabilities, credentials, and security policies. Also new: remote SSH connections to enterprise environments, programmatic access tokens for Enterprise/Business, and HIPAA-compliant CLI/IDE usage. Currently macOS-only for desktop connection; Windows support is coming. [5][6][7]

• Claude Code v2.1.142 shipped with Opus 4.7 as the new Fast mode default and comprehensive agent configuration flags. The --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, and --dangerously-skip-permissions flags fully configure dispatched background sessions. Plugins with root-level SKILL.md are now surfaced as skills. Critical stability fixes address daemon crash-loops after binary upgrades, background sessions disappearing after macOS sleep/wake, and agents crash-looping with Chrome extension connected. Users can set CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1 to pin Fast mode to Opus 4.6. [13]

• GitHub Copilot CLI v1.0.48 released with token-price display in model picker and instruction file fixes. The model picker now shows actual token prices for billing users instead of dot indicators. Instruction files with unquoted glob patterns in applyTo frontmatter now apply correctly. CJK/emoji rendering gaps are fixed, /context shows correct per-model token limits, and skill content no longer injects YAML frontmatter metadata to the model. [14]

• Gemini CLI v0.44.0-nightly.20260514 merged Auto modes into a single unified Auto mode and added RAG snippet debugging. Agent registration now uses first-wins priority with project agents taking precedence. Fixes address OAuth token rotation, sandbox permission issues on NixOS, NO_PROXY for network MCP servers, and Vertex AI preview model access. The release also resolves EISDIR errors during file processing and externalizes https-proxy-agent for proxy support. [15]

• OpenCode v1.14.50 fixed HTTP event stream persistence and improved error messaging for model configuration. The release keeps event streams open after initial connect, returns proper busy errors when sessions are already running, and lets invalid small_model config values fall back gracefully. Markdown rendering for session output was restored by default, and the v2 model/provider listing API now supports instance directory and workspace queries. [16]

• OpenAI Codex shipped four alpha releases (0.131.0-alpha.14 through alpha.17) on May 14 with incremental improvements. These Rust-based releases continue the rapid iteration cadence of approximately one release every 4–8 hours, though no detailed changelogs were provided for individual alpha builds. [17]