Claude Code for Web — Async Cloud Agent #
- Claude Code for web — a new asynchronous coding agent from Anthropic (Simon Willison, Substack) — Willison’s preview notes: Claude Code for web is effectively a sandboxed instance of
claude --dangerously-skip-permissions running in Anthropic’s container infrastructure. Developers access code.claude.com from the browser, describe a task, and the agent works asynchronously — continuing after the browser tab is closed, with tasks persisting across sessions and devices. Key insight: architecturally identical to local Claude Code, but the execution environment is Anthropic’s cloud. The “dangerous permissions” are safe because it’s sandboxed, not because it’s supervised. - Claude Code Routines: How to Run 24/7 AI Agents Without Keeping Your Computer On (MindStudio) — Practical guide to Claude Code Routines: scheduled agents that run on Anthropic’s infrastructure without requiring a local process to stay alive. Use case: persistent agents (nightly analysis, scheduled pipeline refreshes) rather than interactive sessions. Distinguishes from Remote Control (runs on local machine) and from Claude Code for web (single async session). Routines complete the execution matrix: interactive → async session → scheduled.
Skills — Open Standard #
- Claude Skills are awesome, maybe a bigger deal than MCP (Simon Willison, Substack) — Willison’s argument: Skills are conceptually simpler than MCP (a Markdown file describing how to do a task + optional CLI tools) but solve the same integration problem with dramatically lower token overhead. Key point: MCP’s token consumption was a real context-window cost; Skills avoid it because the LLM already knows how to call
cli-tool --help. Anthropic has since turned the skills mechanism into an open standard (agentskills/agentskills GitHub repo), signalling cross-tool portability as the intended trajectory.
Cross-links #
- Vibe Coding Approaches Claude Code for web completes the async execution model — interactive terminal, async cloud session, scheduled Routines — that Karpathy’s “agentic engineering” framing requires for long-running oversight work.
- Claude Integrations The agentskills open standard is the Skills equivalent of MCP’s cross-tool aspiration — if it gains adoption, Skills become a cross-platform agent instruction format.
- Emerging pattern: Anthropic is now shipping three distinct execution modes for Claude Code (local interactive, cloud async, cloud scheduled). Each removes a different friction: latency, machine dependency, session dependency. Convergence point: “Claude as ambient background agent.”
- Keyword suggestion:
"claude code routines" OR "claude code for web" async scheduled — Routines is under-covered relative to the interactive features; practitioner experience pieces will appear in the next cycle.