Lucas Bento

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

Tips for getting the most out of OpenAI's Symphony (what actually worked for me): - Have an agent do an assessment of the task first to pick which model to execute on. You don't need the most powerful model to write docs - The "Merging" lane doesn't need 5.4 to land a PR. Cheaper models handle them just fine and can save you some tokens - Create a Testing lane so lightweight models can test the output and refuse/approve the work before you review it - Have Codex improve on the web dashboard, I had it create a log console per task so I can see which models are being used for what tasks and what is slowing them down - Create helpers for common tasks so it's clear to agents how to get them, e.g. I created a small bash script so agents can easily fetch ticket details from Linear and update the task - After each batch of tasks, spin an agent to analyze past sessions and identify bottlenecks, continuous learning and improvement Symphony is quite impressive but will still eat your token budget alive if you don't pay attention

1 million context window: Now generally available for Claude Opus 4.6 and Claude Sonnet 4.6.

Pierre diffs[dot]com is in everything now every agent tool, every coding workflow, every diff review ever and it's read-only in all of them we made it editable fix the bug right in the diff. no round-trips. (vim mode too, obviously)