Agent builder reports similar task completion with just Bash and a media viewer
The builder describes Bash, a command-line shell, as a routing layer that lets agents combine Git, Python and other tools into one-off workflows.
TLDR
An agent builder says they removed every model-facing tool except Bash and a media viewer when rebuilding their agent setup. They report that task completion stayed in the same range, with fewer tool boundaries during complex work and verification. They argue that models can assemble disposable command-line workflows in seconds, while acknowledging that the models still make mistakes.
Combined views
—
1 Source, first seen ago
