Product agents need workbenches, not tool lists

Brex's spend-audit agent benchmark compares native product tools with a bash-shaped workspace. The signal: complex agents need somewhere to filter, save, script, and inspect intermediate work instead of pushing every raw result back into the model context. Why it matters: For internal product agents, the interface can be the bottleneck. More tools help agents reach systems, but a bounded workspace helps them turn large messy data into useful context without blowing up tokens, latency, or recall.

Try this: When designing a long-running agent, give it safe commands, files, validators, and a submission harness. Keep Docker or stronger isolation for cases that need arbitrary binaries or untrusted execution; do not confuse a bash-shaped interface with a security boundary.