A new research system lets a talking AI assistant hand off hard problems to a coding agent, then automatically patches its own handoff process when something goes wrong.
The paper describes DuplexAgent, a voice-agent design that splits the work between a "full-duplex" model, one that can listen and speak at the same time instead of waiting for turns, and a separate reasoning or coding agent for jobs too complex for live conversation, like research or writing code. Managing that handoff, accepting a task, tracking it, canceling it, swapping in another agent, delivering results without an awkward silence, is handled by a "harness" built from six editable modules rather than one block of hard-coded rules. The paper's real contribution is Duplex-Harness-RSI, a closed loop that runs simulated, timed test conversations, flags which module caused a given failure, and has an LLM called the Harness Editor rewrite just that module. An "Exam Planner" decides what to test next based on past weak spots and an archive of previous fixes.
Most voice-agent demos paper over the seam between talking and thinking. What's interesting here is that the handoff logic itself becomes something that improves from evidence instead of a developer's hunches, using the same reasoning and coding models that serve the user to also serve as the system's own mechanics. If that holds up outside a lab, it is a plausible blueprint for any product pairing a responsive front end with slower specialist back ends, think customer-service bots or pair-programming tools.
The reported gains come from the authors' own benchmarks against their own comparison systems, so treat "outperforms" as a hypothesis, not a verdict, until outside researchers get to test it.