Key Signals

• Anthropic launches Routines for Claude Code, enabling event-driven automation without local infrastructure. Routines allow developers to configure automated coding workflows triggered by schedules, API calls, or GitHub webhooks. This positions Claude Code beyond interactive sessions into persistent, asynchronous agent territory — directly competing with GitHub Copilot's cloud agents and OpenAI Codex workflows. The feature supports use cases like automated issue triage, cross-language SDK synchronization, and CI-failure response, signaling Anthropic's push toward always-on coding agents. [5][6]

• GitHub Copilot CLI v1.0.49 introduces MCP server registry search, expanding the tool ecosystem. The new experimental /mcp search command lets developers discover and install MCP servers directly from a registry, alongside tool search with deferred loading for MCP and external tools. This lowers the barrier to extending Copilot CLI with third-party integrations and signals GitHub's commitment to MCP as the standard protocol for agent-tool interoperability. [1][2]

• Copilot Memory now stores user-level preferences that follow developers across all repositories. Previously limited to repository-scoped information, Copilot Memory can now capture personal preferences — commit style, PR structure, communication tone — in early access for Pro and Pro+ users. This cross-repo personalization layer represents a shift toward long-term developer relationship modeling, making Copilot increasingly context-aware without per-project configuration. [3]

• AWS embeds formal automated reasoning into Kiro's requirements phase, catching defects before code generation begins. Kiro's new Requirements Analysis feature uses SMT solvers to mathematically prove contradictions and gaps in specifications — not probabilistic flagging, but formal verification. In internal testing across 35 projects with 1,400+ acceptance criteria, roughly 60% of first-draft requirements needed refinement. This neurosymbolic approach differentiates Kiro from competitors relying solely on LLMs for code quality. [11]

• Block transfers Goose to the Linux Foundation's Agentic AI Foundation, resolving governance barriers to enterprise adoption. Block's trademark ownership over Goose was hampering enterprise uptake despite rapid community adoption. The Agentic AI Foundation, alongside MCP and Agents.MD, now provides neutral governance. This move legitimizes Goose as a vendor-neutral agentic coding tool and strengthens the Linux Foundation's emerging role in AI agent infrastructure standards. [10]

• GitHub deprecates Grok Code Fast 1 across all Copilot experiences, recommending GPT-5 mini and Claude Haiku 4.5. The deprecation affects Copilot Chat, inline edits, ask/agent modes, and code completions. Enterprise administrators need to update model policies. This reflects the rapid model churn in AI coding tools and GitHub's multi-model strategy — cycling through providers as newer models outperform older ones. [4]

AI Coding News

• Block handed its AI coding agent Goose to the Linux Foundation's new Agentic AI Foundation to solve enterprise governance concerns. After open-sourcing Goose and seeing rapid adoption, Block encountered "headwinds" because it still owned the trademarks, creating uncertainty for enterprise users. The newly formed AAIF, co-founded by Block's former head of open source Manik Surtani, now hosts Goose alongside MCP and Agents.MD as core projects. The foundation structure provides vendor-neutral governance with both corporate and developer buy-in, including AWS involvement. [10]

• AWS launched Requirements Analysis in Kiro, using a 50-year-old formal logic engine to catch bugs in software specifications before code is written. The feature combines LLM rewriting of vague requirements into testable criteria with SMT solver-based mathematical proofs that identify contradictions, ambiguities, and gaps. Findings surface as plain-language questions resolvable in 10–15 seconds each. Kiro also shipped Parallel Task Execution (cutting implementation time by ~75% for large specs) and Quick Plan for single-pass requirement generation. AWS announced Shawn Bice as new VP of AI Services leading the Automated Reasoning Group, with VP Swami Sivasubramanian stating they are "at an inflection point with Agentic AI." [11]

Feature Update

• GitHub Copilot CLI v1.0.49-0 adds experimental MCP server registry search, tool discovery with deferred loading, and granular memory scoping. The release introduces /mcp search to discover and install MCP servers from the registry, a "None" reasoning effort option, and COPILOT_PLUGIN_DIR_ONLY for deterministic plugin sets. Memory permission prompts now display scope visibility. Fixes address hook firing for sub-agent tool calls, tmux rendering in ghostty/WezTerm/kitty terminals, and MCP OAuth token scoping. [1]

• GitHub Copilot CLI v1.0.49-1 patches prompt mode to auto-load workspace MCP sources when the folder is already trusted. This removes the manual step of enabling MCP sources in prompt mode for previously trusted workspaces, streamlining the developer workflow when switching between interactive and prompt modes. [2]

• GitHub Copilot Memory expands to user-level preferences in early access for Pro and Pro+ users. Copilot can now store and apply personal preferences — commit style, PR structure, communication tone — that persist across all repositories and Copilot agents without affecting other users. User memories are manageable at github.com/settings/copilot/memory. This complements the existing repository-level memory with portable personalization. [3]

• Grok Code Fast 1 is deprecated across all GitHub Copilot experiences effective May 15, 2026. The model is removed from Copilot Chat, inline edits, ask/agent modes, and code completions. Suggested alternatives are GPT-5 mini and Claude Haiku 4.5. Enterprise administrators should verify model policy settings and enable access to alternatives through Copilot settings. No action is needed to remove the deprecated model. [4]

• Claude Code v2.1.143 ships plugin dependency enforcement, expanded claude agents CLI flags, and extensive background session stability fixes. New capabilities include projected context cost in the plugin marketplace, worktree.bgIsolation: "none" for repos where worktrees are impractical, and PowerShell -ExecutionPolicy Bypass by default on Windows. The claude agents command now accepts --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, and --effort flags. Key fixes address background session preservation across sleep/wake, macOS file permission errors, Windows Terminal responsiveness, and daemon spawn fallbacks. [5]

• Anthropic introduces Routines for Claude Code, enabling scheduled, API-triggered, and webhook-based automated coding workflows. Routines run on Claude Code's cloud infrastructure without requiring local cron jobs or servers. Scheduled routines handle recurring tasks like bug triage and documentation drift scanning. API-triggered routines expose authenticated endpoints for integration with deployment pipelines and monitoring. GitHub webhook routines auto-launch sessions on PR events, tracking CI failures and responding to comments across the PR lifecycle. [6]

• Gemini CLI v0.44.0 nightly build adds RAG snippet debugging, enterprise gateway auth fixes, and NixOS sandbox support. The release exposes RAG snippets to local log files for debugging, prevents conflicting credentials on enterprise gateways with native optional API key support, and respects NO_PROXY for network-based MCP servers. Additional fixes address Vertex AI preview model access, sandbox permission errors on NixOS, and proxy support via externalized https-proxy-agent. [7]

• OpenCode v1.15.0 introduces an Effect-based core event system for complete event delivery across sessions and integrations. The release fixes handling of invalid custom tool module exports, project instruction lookup errors, and versioned event projector lookups. Desktop gains auto-hide menu bar on Linux and Windows. The JavaScript SDK restores previously missing session and message event types. [8]

• OpenCode v1.14.51 adds experimental background subagents and updates LiteLLM compatibility for GPT-5 support. Background subagents let tasks continue running while the developer works on other things. The release requires LiteLLM v1.85.0-rc.2+ for current GPT-5 and tool-call behavior, fixes Azure gpt-5.5 requests through the completions API, and adds MCP connection status with direct re-auth flows in the desktop app. Seven community contributors participated in this release. [9]

• OpenAI Codex published four Rust CLI alpha releases (0.131.0-alpha.18 through alpha.21) on May 15, indicating rapid iteration on the rewrite. The releases had no detailed changelogs, but the cadence of four builds in a single day signals active development momentum on the Rust-based Codex CLI. The project continues its transition from the original TypeScript implementation. [12]