First implication: teams should treat permission design and review of agent changes as workflow decisions, not assume the approval toggle isolates execution. The source establishes documented behavior and caveats, not independent security testing or measured productivity gains.
Thesis
Mirror CLI makes the coding-agent harness, rather than the model alone, the operational surface developers must understand.
Observed facts
Reflection’s developer documentation identifies Mirror as a terminal coding agent in early beta.
1The documentation says Mirror uses Reflection and Beam-1.0-501B-A23B by default.
1Manual mode pauses before shell commands, file edits, connected-service tools, and most extension tools.
1Reflection warns that approval mode is not a sandbox and approved tools run with the user’s account permissions.
1
Mechanism
The approval setting governs when Mirror pauses for user review, while tools execute with the permissions of the user’s machine account.
Why now
the public beta makes these operating boundaries concrete for teams deciding how to trial an agent in real project directories.
Implications
teams should treat permission design and review of agent changes as workflow decisions, not assume the approval toggle isolates execution.
Evidence boundary
this assessment relies on Reflection’s product documentation, which says commands and behavior may change during early beta.
Watch conditions
whether later releases add a sandbox boundary or otherwise change how tools inherit local permissions would materially alter this assessment.