I've been using Seqera AI CLI in the past couple of days until I exhausted the free credits :P, and I thought I might share a couple of immediate feedback in case it's useful. context: used it on MacBook OS 26.3 native terminal (set to bash) using a conda environment to keep my env tidy. (1) while using it, it stops scrolling and I have to scroll manually to keep reading messages. this is moderately annoying, but not a blocker. (2) plenty of requests for permissions, but mostly for running GitHub commands and not about the files it decides to edit: I've ended up realising it modified for reasons I didn't agree with logic of some modules or scripts which had nothing to do with the debugging tasks I set it out to perform. This is a little dangerous, because on one side it bugs you with plenty of command line requests, but at the same time they are general. both Claude Code and openAI codex as well as Antigravity enter a planning mode before execution, and therefore it is possible to see and discuss and edit the logic of the changes to be introduced. SeqeraAI attempts to be very autonomous in decisions which should not be: a plan should be presented, approved and once it is approved then user could also leave the screen rather than approving every git command or ls command the agent wants to run. My two cents. Happy to provide more details if requested.