I tested Claude's new agent features for a day. Cowork, Dispatch, computer use, Claude Code in the desktop app. All of it.
My honest take: Anthropic is getting close. Not there yet, but close. And the direction they're going is exactly right.
I built a custom agent system that's been handling automation for months, so I tested these tools against what I've already learned works and what breaks.
Claude Code Desktop's visual diff reviewer cuts code review time in half. Inline comments, worktree isolation for parallel sessions, and a live browser preview that actually works without thrashing.
Cowork's connector catalog (50+ integrations—Slack, Gmail, Jira, Notion, Google Calendar) handles task automation that would take weeks to script. The catch: it forgets everything between sessions, so it can't build on past decisions.
Computer Use's screen automation is honest-to-god research preview. It sees what's on screen and can click/type, but hits a wall at 50% reliability. Useful for one-off tasks, dangerous for critical workflows.
Dispatch's cross-platform execution (mobile task assignment routed to Claude Code via Slack/Discord/Telegram) is the piece that actually feels new. Turns your phone into a command center for desktop automation.
The biggest insight isn't that these tools work—it's that three major companies (Anthropic, OpenAI, Google) shipped nearly identical "agent on your desktop" products within two weeks. That convergence is validation that someone figured out the right problem. But persistent memory, rate limit impacts on production, and vendor lock-in are still unresolved.
The part that surprised me most is in the full post.
Read the full breakdown: https://thoughts.jock.pl/p/claude-cowork-dispatch-computer-use-honest-agent-review-2026
Subscribe for weekly posts: https://thoughts.jock.pl
Originally published on Substack. View on Substack
Top comments (0)