Key Signals

• Anthropic launches Claude Code Routines, turning its coding agent into a persistent background worker. Routines let developers configure a prompt, repository, and connectors once, then trigger execution on a cron schedule, via an API call, or in response to GitHub webhook events. Unlike the earlier /schedule command, routines run on Anthropic's cloud infrastructure with no interactive approval prompts — shell commands, skills, connectors, and MCP servers are all available autonomously. Pro subscribers get 5 routine runs per day, Max users 15, and Team/Enterprise plans 25, making this the most significant step yet toward always-on AI coding assistants. [1]

• Claude Code's redesigned desktop app introduces an integrated terminal, side chats, and rearrangeable panes. The rebuilt interface positions users as "orchestrators" managing multiple agentic sessions across repositories simultaneously. A new side-chat feature lets developers ask questions without interrupting an active agent, and the app now includes a terminal, improved diff viewer for large changesets, an in-app file editor, and HTML/PDF preview — closing a key gap with OpenAI's Codex desktop experience. Anthropic also rebuilt the app internals for streaming performance. [2]

• GitHub now offers model selection for Claude and Codex third-party coding agents on github.com. Users can choose between Claude Sonnet 4.5, Claude Sonnet 4.6, Claude Opus 4.5, and Claude Opus 4.6 for the Anthropic agent, or GPT-5.2-Codex, GPT-5.3-Codex, and GPT-5.4 for the OpenAI agent. This gives developers fine-grained control over cost, latency, and capability trade-offs directly within their existing Copilot subscription, and mirrors the model picker already available for GitHub's own Copilot cloud agent. [3]

• UK AI Security Institute evaluation reveals Claude Mythos Preview as the first model to autonomously complete a full 32-step corporate network takeover. In the "The Last Ones" simulation — covering reconnaissance through total compromise — Mythos Preview succeeded end-to-end in 3 of 10 attempts and averaged 22 of 32 steps, far outpacing Claude Opus 4.6's average of 16. It also completed expert-level capture-the-flag tasks 73% of the time, a benchmark no model could reach before April 2025. Access remains restricted to roughly 40 organizations via Anthropic's Project Glasswing, underscoring the dual-use implications these capabilities carry for both defensive tooling and offensive risk. [4]

• Gemini CLI ships v0.38.0 stable and v0.39.0-preview.0 on the same day, adding memory skills, context compression, and a unified subagent tool. The stable v0.38.0 introduces a background memory service for automatic skill extraction, a ContextCompressionService for longer sessions, TerminalBuffer mode to eliminate UI flicker, and background process monitoring. The preview v0.39.0-preview.0 builds on that with a /memory inbox command for reviewing extracted skills, JSONL streaming for chat recording, persistent policy approvals, and a refactored invoke_subagent tool that consolidates legacy subagent wrappers. [6][7]

• Copilot CLI v1.0.26 fixes a critical context-compaction crash and tightens security. Agent sessions previously could fail with unrecoverable errors when context compaction split a tool call across a checkpoint boundary — now resolved. The release also binds the ACP server to localhost only, adds plugin hook environment variables, renames "steering" to "remote control" throughout the CLI, and consolidates duplicate instruction files to reduce per-turn token waste. [5]

AI Coding News

• Spring Framework creator Rod Johnson unveiled Embabel, a JVM-based agentic AI framework that uses game-AI pathfinding instead of LLM-driven routing. Embabel, built on Spring Boot and written in Kotlin, applies Goal-Oriented Action Planning — a deterministic algorithm from game development — to dynamically compose execution paths through strongly typed Java methods. When a field carries a Jakarta EE validation constraint, the framework surfaces it to the LLM as part of the prompt and loops back on failure, making the model a participant in Java's type system rather than an opaque endpoint. Embabel supports per-step model selection across OpenAI, Anthropic, and Llama providers and offers direct integration points for Claude Code and GitHub Copilot. [10]

• OpenAI expanded its Trusted Access for Cyber program with GPT-5.4-Cyber, a specialized model for vetted cybersecurity defenders. The program provides approved organizations with access to purpose-built models designed for cyber defense workflows, alongside strengthened safeguards as AI cybersecurity capabilities advance. This launch parallels Anthropic's restricted-access Project Glasswing for Claude Mythos Preview, signaling an emerging pattern where frontier AI labs gate their most capable security-related models behind vetted-access programs. [11]

Feature Update

• Claude Code v2.1.108 adds a session recap feature, 1-hour prompt caching, and slash-command discovery via the Skill tool. The new ENABLE_PROMPT_CACHING_1H environment variable opts into a 1-hour prompt cache TTL across API key, Bedrock, Vertex, and Foundry providers. A recap feature provides context summaries when returning to a session after being away. The model can now discover and invoke built-in commands like /init, /review, and /security-review through the Skill tool. Additional improvements include better error message differentiation between server rate limits and plan usage limits, reduced memory footprint via on-demand language grammar loading, and 15+ bug fixes covering paste in /login, session title handling, diacritical mark rendering, and plugin auto-update. [8]

• Claude Code v2.1.107 ships a minor improvement: thinking hints now appear sooner during long operations. This quality-of-life change reduces perceived latency by giving users earlier visibility into the model's reasoning process during extended tool calls. [9]

• Claude Code Routines launch as a new automation primitive with three trigger types. Routines are configured once — with a prompt, repository, and connectors — then fired on a schedule, from an API call, or via GitHub webhooks. Existing /schedule tasks automatically convert to routines. They run on Anthropic's cloud infrastructure autonomously, with no permission prompts, and can execute shell commands, invoke skills, and call MCP servers. Daily run limits are tiered: 5 for Pro ($20/month), 15 for Max ($100–$200/month), and 25 for Team/Enterprise plans. [1]

• Claude Code desktop app receives a major redesign centered on multi-session orchestration. The new layout surfaces all active and recent agent sessions, adds context-aware side chats, and introduces an integrated terminal, rebuilt diff viewer, in-app file editor, and HTML/PDF preview panes — all of which can be rearranged freely. The app's chat/Cowork/Code switcher moves from a top modal to sidebar icons, and the entire rendering pipeline was rebuilt for streaming performance. [2]

• GitHub enables model selection for Claude and Codex third-party coding agents. Claude users can now pick from Claude Sonnet 4.6, Claude Opus 4.6, Claude Sonnet 4.5, and Claude Opus 4.5; Codex users can choose GPT-5.2-Codex, GPT-5.3-Codex, or GPT-5.4. Access is bundled with existing Copilot subscriptions; Business/Enterprise administrators must enable the relevant agent policy under Settings > Copilot > Cloud agent. [3]

• Copilot CLI v1.0.26 delivers 18 fixes and improvements spanning agent stability, security, and UX. Beyond the critical context-compaction fix, the release stops spurious directory-access prompts inside find -exec blocks, ensures Anthropic BYOM correctly includes image data, makes the Remote tab show Copilot coding agent tasks without requiring a pull request, adds keyboard-navigable session scope selectors, and consolidates instruction files with applyTo patterns into a table to reduce context-window usage. [5]

• Copilot CLI v1.0.26-0 pre-release ships a subset of the v1.0.26 changes. This pre-release includes the plugin hook environment variables, session scope selector improvements, instruction file consolidation, ACP localhost binding, marketplace URL parsing fix, enterprise hostname handling, Windows LSP file URI paths, and relative path resolution. [12]

• Gemini CLI v0.38.0 introduces background memory, context compression, and TerminalBuffer mode. The stable release adds a background memory service for skill extraction, a ContextCompressionService for managing long sessions, TerminalBuffer mode to solve UI flicker, background process monitoring and inspection tools, browser agent improvements, an experimental adk.agentSessionNoninteractiveEnabled setting, and a fix for Linux sandbox ARG_MAX crashes. [6]

• Gemini CLI v0.39.0-preview.0 previews a /memory inbox command and unified subagent architecture. Users can now review and manage skills extracted by the background memory service through a dedicated inbox. The release also refactors subagent invocation into a single invoke_subagent tool, migrates chat recording to JSONL streaming, implements context-aware persistent policy approvals, and adds an useAgentStream hook for real-time agent output rendering. [7]

• Gemini CLI v0.39.0-nightly.20260414 adds skill patching, Plan Mode routing fallback, and decoupled ContextManager. The nightly build introduces skill patching support with /memory inbox integration, a silent fallback for Plan Mode model routing, a new decoupled ContextManager and Sidecar architecture, a tool-controlled display protocol, mouse-click cursor positioning in AskUser prompts, and auth block support for MCP servers in agent configurations. [13]

• OpenAI Codex CLI ships four Rust alpha builds (0.121.0-alpha.7 through alpha.10) in a single day. These automated alpha releases of the Rust-based Codex CLI continue rapid iteration on the rewrite, though no individual changelogs were published for these builds. The cadence — four releases in approximately 16 hours — indicates active development toward the next stable milestone. [14][15][16][17]