Anthropic just dropped Claude Duet. No fanfare. No blog post tsunami. Just a waitlist flip and a handful of devs waking up to access emails at 3 a.m.
The app sits inside your IDE. VS Code. Cursor. Windsurf. It doesn't answer questions. It does the work.
You type: "Refactor the auth module to use Passkeys. Update tests. Open a PR." Duet reads the repo. Maps dependencies. Edits twelve files. Runs the suite. Pushes the branch. Opens the PR with a description that actually makes sense.
Thirty-seven minutes. Zero human keystrokes after the prompt.
That's the claim. Early access folks are posting screen recordings. They look real.
Here's what's different. Previous agent tools — Devin, Cursor's agent, even Claude's own computer use — they needed babysitting. You had to approve every file write. Every terminal command. Every git action.
Duet asks once. "Proceed with full autonomy?" You say yes. It vanishes into the codebase.
The trick is a new planning layer Anthropic calls "task graphs." Not a linear todo list. A directed acyclic graph of subtasks with explicit success criteria. Each node validates itself before the next fires. If test coverage drops, the graph rewinds. If a dependency conflict pops, it forks a resolution branch.
You watch a live DAG visualization. Nodes green. Nodes yellow. Nodes red — then auto-healing green again.
It's mesmerizing. Also slightly terrifying.
Pricing isn't public. Rumor says seat-based plus compute. Enterprise deals already signed — three Fortune 100 names whispered in Discord.
Anthropic's bet: agents aren't a feature. They're the product. The chat interface becomes legacy UI.
Competitors have weeks. Maybe days.
What happens when the waitlist opens to everyone?

Memuat komentar...