Hey, this is a known issue. There can be a mismatch between the mode shown in the UI and what the model actually receives in the system prompt. The bug shows up in both the IDE and the CLI.
A working workaround is to always start a new chat in Agent mode, and don’t switch modes within the same chat. If you need to keep context, you can also try adding this to your prompt: You are in Agent mode. Execute commands directly. This sometimes helps switch the behavior back.
Related reports for context:
- GPT-5.4 thinks its in ask mode even in agent mode GPT-5.4 thinks its in ask mode even in agent mode
- Agent mode thinks its in Ask mode Agent mode thinks its in Ask mode
The team is aware of the issue. There’s no timeline yet, but each report helps with prioritization.
Let me know if the workaround doesn’t help.