Twin Agent makes the hint budget the security knob.

A July 21 paper pairs two near-identical agents: an Explore Agent that inspects untrusted input but holds no tools, and a Safe Agent that takes privileged actions using its own context plus length-constrained hints from the explorer (arXiv 2607.19595). Borrowing from residual coding, longer hints raise both task utility and injection risk, so you tune the tradeoff explicitly rather than discovering it in production. It beat both undefended agents and prior privilege-separation baselines on the utility/security frontier across SWE-bench Lite and AgentDojo.