me

@ QIR

me

Working @ QIR

  • home
  • about
  • projects
  • stacks
  • posts
  • contact
  • experimental
    • Server Action
    • Date Picker
    • Debounce
    • Recursive Tree

Made with Next.js, Tailwind, Sanity.

AI
LLM
Dev Tools
AI Coding

18 April 2026

Scheduled agents and remote CLIs for Claude and Copilot

Ryan Knight

Ryan Knight

what new in ai cover

The week of April 13, 2026 brought developer-facing releases centered on moving agent work off the local machine and letting running sessions be supervised from elsewhere. Anthropic opened a research preview of Claude Routines, which runs saved Claude Code configurations on managed cloud infrastructure, and rebuilt the Claude Code desktop application around parallel sessions. GitHub placed Remote Control for Copilot CLI sessions into public preview, allowing an in-flight session to be steered from the web and from a phone. Vercel added Claude Opus 4.7 and Qwen 3.6 Plus to the AI Gateway, broadening the set of models that applications can target through a single endpoint. Together these releases change what developers can hand off, resume, or observe without keeping a terminal in the foreground.

Claude Routines: scheduled agents on Anthropic's cloud

Anthropic opened a research preview of Claude Routines on April 14, 2026, a service that runs Claude Code configurations on infrastructure managed by Anthropic (SiliconANGLE; 9to5Mac).

A routine packages a Claude Code prompt, one or more repositories, and a set of connectors into a reusable configuration that executes on Anthropic's cloud rather than on a developer's workstation, so runs continue after the originating laptop is closed. Triggers include fixed cadences such as hourly, daily, weekdays, or weekly, GitHub events such as pull requests or releases, and direct API calls. The preview is available to Claude Code Pro, Max, Team, and Enterprise subscribers, with daily run caps reported as five for Pro, fifteen for Max, and twenty-five for Team and Enterprise; each routine requires a web-enabled model (SiliconANGLE; 9to5Mac).

The scheduled and event-triggered model enables applications that previously required a purpose-built worker. A routine bound to pull-request events can scan CI output, identify failing steps, and post a summary comment on the PR. A daily routine can consolidate dependency updates across several repositories into one draft PR. Each of these patterns now ships as a saved configuration rather than a bespoke CI job.

A rebuilt Claude Code desktop app

Anthropic also rebuilt the Claude Code desktop app around parallel sessions, an integrated terminal, an in-app file editor, and a rebuilt diff viewer aimed at large changesets (MacRumors; VentureBeat).

Parallel sessions surface in a new sidebar so multiple Claude Code tasks run side by side within a single application window, each retaining its own context. The integrated terminal runs tests and builds from inside the app, and the in-app file editor supports spot edits without switching to an external editor. The diff viewer targets large, multi-file changesets specifically, which is the shape of change that agent-driven edits tend to produce (MacRumors; VentureBeat).

The redesign affects workflows that run more than one coding task at a time. A long-running refactor can sit in one pane while an unrelated feature is iterated on in another, without one task blocking the other. Teams reviewing an agent-produced multi-file change can inspect it in the rebuilt diff viewer and run builds in the same window. The combination reduces the number of separate windows required to operate Claude Code across a typical feature branch.

Copilot CLI Remote Control on web and mobile

GitHub moved Remote Control for Copilot CLI sessions into public preview in a changelog entry dated April 13, 2026, adding the ability to monitor and steer a running CLI session from github.com and from mobile (GitHub Changelog, April 13 2026).

A session started on a local workstation can now be observed, paused, instructed, and resumed from a browser tab or a phone, so the developer who initiated the run does not have to remain at the originating terminal for its full lifetime. The same release series added a --list-env flag for prompt mode, a /statusline command for customizing the status bar, Remote MCP server config defaults, and Claude Opus 4.7 as a selectable model inside the Copilot CLI (Releasebot, Copilot CLI).

The capability fits several concrete patterns. A long-running refactor started before a commute can be checked and redirected from a phone without returning to the workstation. An on-call engineer can intervene in a paused Copilot CLI session from a browser on a different machine. The Remote MCP config defaults also reduce per-project boilerplate when connecting Copilot CLI to a team's own MCP servers.

Claude Opus 4.7 and Qwen 3.6 Plus on the Vercel AI Gateway

Vercel's AI Gateway added Claude Opus 4.7 and Qwen 3.6 Plus during the week's release cycle, alongside Bytedance's Seedance 2.0 video generation model (Vercel Weekly, April 13 2026).

Claude Opus 4.7 is routed through the Gateway as a model identifier applications can select, with Vercel describing it as optimized for long-running asynchronous agents and multi-step tasks. Qwen 3.6 Plus joins the Gateway with a 1-million-token context window and Alibaba-reported improvements on tool calling, long-horizon planning, and multilingual work. Both are reachable through the same endpoints that the Gateway already exposes for other hosted models (Vercel Weekly, April 13 2026).

The addition changes what Gateway-backed applications can do without integration work. A single endpoint can A/B a long-context summarization pipeline against Qwen 3.6 Plus's 1M-token window alongside a smaller model, with no new SDK integration. An application already calling a prior Opus tier can move production traffic to Opus 4.7 by swapping the model identifier. The Seedance 2.0 addition extends the same endpoint pattern to generated-video workflows that were previously handled outside the Gateway.

This week's releases touch three different parts of the agent workflow: where the work runs, where it is supervised, and which models it can address. Claude Routines moves scheduled and event-driven agent work onto Anthropic's infrastructure; Copilot CLI Remote Control and the redesigned Claude Code desktop app change what a session looks like while it is in progress; Vercel's Gateway additions broaden the set of models reachable behind a single endpoint. The common thread is a reduction in the number of always-on local processes required to run and observe autonomous coding tasks.

Sources

  • https://siliconangle.com/2026/04/14/anthropics-claude-code-gets-automated-routines-desktop-makeover/
  • https://9to5mac.com/2026/04/14/anthropic-adds-repeatable-routines-feature-to-claude-code-heres-how-it-works/
  • https://www.macrumors.com/2026/04/15/anthropic-rebuilds-claude-code-desktop-app/
  • https://venturebeat.com/orchestration/we-tested-anthropics-redesigned-claude-code-desktop-app-and-routines-heres-what-enterprises-should-know
  • https://github.blog/changelog/2026-04-13-remote-control-cli-sessions-on-web-and-mobile-in-public-preview/
  • https://releasebot.io/updates/github/copilot-cli
  • https://releasebot.io/updates/anthropic
  • https://community.vercel.com/t/vercel-weekly-2026-04-13/38153