Full‑duplex voice agents are converging on a collaboration pattern where a live channel provides continuous listening and speaking, while search, reasoning and coding are handled via asynchronous delegation. The duplex model supports uninterrupted interaction but struggles with complex reasoning and tool use; conversely, coding agents can plan and execute long‑running tasks but their sequential interface does not fit live conversation. A harness is needed to coordinate task acceptance, progress, cancellation, replacement and result delivery while keeping the dialogue responsive. Existing harnesses rely on tightly coupled heuristics, making systematic evidence‑driven improvement difficult. DuplexAgent expresses the workflow as six editable modules and introduces Duplex‑Harness‑RSI, a closed‑loop recursive self‑improvement system:
- A simulator automatically creates timed test conversations, runs the system and produces failure traces that pinpoint modules needing repair.
- An exam planner selects the next set of tests based on observed weaknesses and the repair archive.
- A harness editor, powered by reasoning LLMs and coding agents, proposes targeted module changes.
These components form a self‑improving loop where the capabilities that serve the user also refine the system’s coordination. Experiments on intelligence, agentic and duplex benchmarks show that DuplexAgent combines continuous interaction with demanding reasoning and complex task execution, achieving higher spoken‑knowledge and executable‑tool scores than comparable delegated systems while preserving strong interruption response. An ablation of the harness demonstrates that the modular, verifiable loop with diagnosis and repair archive outperforms the initial harness and repeated editing without such support.
Review