Wasting tokens to have subagents “summarize” what would have been directly ingested in the previous version of AG is not an improvement. It is a waste of tokens that ends in an objectively worse result. This isn’t 2023, context windows are not so small that “summarizing” context is any longer necessary.
Overall I’m not sold on “subagents” at all. Every tool I used that leverages them just seems to waste tokens and leads to a worse result. Theoretically, there might be a way to pull it off while avoiding the major issues (including exponential token burn), but it doesn’t appear any attempt has been made to solve the hard problems as much as existing (bad) subagent patterns have just been copy/pasted into antigravity.
The only consolation is that I can continue to use AG IDE, which doesn’t appear to use the subagent madness. I would really like to use the cli, tbh, but it doesn’t support workflows it seems, and also doesn’t have a way to turn off subagents. If you could just provide a more straight-forward subagentless mode in the CLI, I’d be much happier.
I only use AG because it’s currently the best subscription deal for serious work (so I can use a bit of Opus and a lot of Gemini), however that’s getting worse and worse with all these “updates” (how about updating to Claude 4.7?), to the point I might have to scan the room for competition again.