Executor overhauls tool approvals with four modes

The binary choice between complete agent autonomy and tedious human-in-the-loop confirmation is no longer workable for production workflows. Executor's new four-mode approval matrix directly addresses this tension by decoupling token permissions from human authorization channels.

• –Browser isolation eliminates prompt injection escape hatches by ensuring tokens granted to background agents cannot authorize their own paused actions.
• –Tiered elicitation accommodates everything from fast, low-risk local prototyping to headless, production-grade agent services.
• –Decoupling harness prompts from external browser dashboards gives developers architectural flexibility across CLI and hosted setups.
• –Paves the way for incoming intent-based permissioning, aligning tool authorization with semantic agent intent rather than static endpoint checks.