Auset Brain

claude-code-massive-update-march-2026

3 min read

Claude Code Massive Update — March 2026

New Features

1. Computer Use (Mac OS — Research Preview)

  • Claude points, clicks, navigates the computer like a real user
  • Opens files, browses web, interacts with dev tools — zero setup
  • Always asks permission before acting
  • Built-in prompt injection safeguards
  • Mac Pro users only currently

2. Dispatch — Remote Control Claude

  • Assign tasks from phone or anywhere — Claude executes in background
  • Combine with skills.md for repeatable capabilities across entire system
  • Demo: Daniel Sand controlled computer from phone, opened Twitter, liked a post

3. /schedule Command (NEW — Big One)

  • Create recurring cloud-based jobs from terminal
  • Runs even after laptop closed — persistent background worker
  • Anthropic uses internally: auto-resolve CI failures, push doc updates, handle repetitive dev tasks
  • Example: Scheduled job maintains Go twin library of active Python library — auto-syncs

4. Loops Extended to 7 Days

  • Previously 3 days, now 7 for longer-running workflows

5. Effort Levels in Skills/Slash Commands

  • Set thinking effort per-command — overrides session default
  • Low for quick tasks, high for coding/debugging/planning

6. DOM Element Selection (Desktop)

  • Click to select DOM elements instead of describing them
  • Gets: HTML tag, classes, styles, surrounding context, cropped screenshot
  • React: gets source file, component name, props

7. Performance — 65% Faster

  • SSR → static architecture (Vite + TanStack Router, edge-served)
  • TTFB down 65%, components 50% faster

8. Improved /init Command

  • Interviews you to set up Claude Code in repo (skills, hooks, configs)
  • Enable: CLAUDE_CODE_NEW_INITIALIZATION=1 claude

9. Projects in Co-Work

  • Built-in memory across tasks, shared folder + connection setup

Impact on Auset Platform

What Validates Our Approach

  • Dispatch = our agent swarm pattern — We’ve been doing this since March 1. Anthropic is now building it natively.
  • Computer Use = our Chrome MCP + Playwright — We already have browser automation. Native computer use goes further.
  • Schedule = our cron jobs — haiku-dispatcher, sonnet-planner, pr-reviewer all run on cron. /schedule is Anthropic’s version.
  • Projects = our vault/memory system — We built auset-brain vault, MEMORY.md, session-tracker.md. They’re adding shared memory.

What We Should Adopt

  1. Computer Use — Replace Chrome MCP with native computer use when stable. More capable.
  2. /schedule — Could replace our custom cron scripts (haiku-dispatcher.js, sonnet-planner.js) with native scheduling.
  3. Effort levels in skills — Add effort: high to complex skills like /gran, effort: low to simple ones like /progress.
  4. DOM selection — Direct relevance for Katherine (frontend) and Alma (mockup conversion).

What We’re Already Ahead On

  • Multi-agent orchestration — They have dispatch for 1 agent. We have 53 named agents across 9 projects.
  • Voice — They don’t have voice. Clara Desktop + MCP voice channel is ahead.
  • Vault/brain — Their “Projects” memory is per-project. Our vault is cross-project, cross-machine, S3-backed.
  • Named identity — Every agent has a name, a history, a role. Their dispatch is anonymous.

Action Items

  • Test /schedule for CI/CD automation (replace custom crons)
  • Test Computer Use for browser testing (replace Chrome MCP for some workflows)
  • Add effort levels to all skills/commands
  • Evaluate DOM selection for frontend workflow improvement
  • Update boilerplate to leverage new Claude Code features

Graph View