Run Any Agent Harness Under One Interface
Deepgram has introduced Flux TTS, a voice agent technology designed for continuous, context-aware conversations. Unlike standard text-to-speech systems, Flux TTS maintains the pacing and delivery style of previous responses within a streaming session to ensure vocal consistency. The system reports a low latency of 80 ms for speech streaming, making it suitable for real-time interactions. This approach allows the agent to handle interruptions and follow-up questions more naturally by integrating with a Large Language Model that tracks the conversation logic. The Unified Harness Protocol (UHP) establishes a standardized HTTP contract for interacting with AI agent harnesses, decoupling product backends from specific runtime implementations. By providing a consistent interface for task lifecycle management, sessions, and file handling, UHP allows developers to switch between different agent runtimes like Codex and Claude Code without rewriting core application logic. HarnessRouter serves as an open-source implementation of this protocol, enabling teams to route tasks to various harnesses while maintaining a unified event stream and execution environment. This standardization facilitates easier performance comparison, provider flexibility, and robust security through isolated session workspaces.
閱讀原文 ↗A voice agent you can keep talking to
Deepgram has introduced Flux TTS, a voice agent technology designed for continuous, context-aware conversations. Unlike standard text-to-speech systems, Flux TTS maintains the pacing and delivery style of previous responses within a streaming session to ensure vocal consistency. The system reports a low latency of 80 ms for speech streaming, making it suitable for real-time interactions. This approach allows the agent to handle interruptions and follow-up questions more naturally by integrating with a Large Language Model that tracks the conversation logic.
- Flux TTS preserves context from earlier replies within the same streaming session to ensure consistent delivery.
- The system achieves speech streaming latency as low as 80 ms.
- Vocal pacing and delivery are managed by Flux TTS while an LLM tracks the conversation logic.
- The technology is specifically designed to handle interruptions and follow-up questions in voice-based interactions.
- Deepgram demonstrated the tool's capabilities using a car dealership agent scenario discussing hybrid and electric vehicles.
Run any agent harness under one interface
The Unified Harness Protocol (UHP) establishes a standardized HTTP contract for interacting with AI agent harnesses, decoupling product backends from specific runtime implementations. By providing a consistent interface for task lifecycle management, sessions, and file handling, UHP allows developers to switch between different agent runtimes like Codex and Claude Code without rewriting core application logic. HarnessRouter serves as an open-source implementation of this protocol, enabling teams to route tasks to various harnesses while maintaining a unified event stream and execution environment. This standardization facilitates easier performance comparison, provider flexibility, and robust security through isolated session workspaces.
- Unified Harness Protocol (UHP) standardizes the task lifecycle, progress tracking, and session management for AI agent runtimes.
- Harness routing selects the entire agent control loop, whereas model routing only selects the inference endpoint.
- HarnessRouter is an Apache 2.0 licensed, self-hosted UHP server that manages multiple agent backends through a single API.
- The UHP API is modeled after the OpenAI Responses API and uses Server-Sent Events (SSE) for real-time task updates.
- Sessions in UHP provide conversation state and working directory continuity, allowing for follow-up tasks via previous_response_id.
- HarnessRouter ensures security by running sessions in isolated workspaces using separate operating-system users.
- The protocol supports file management where small files are sent in requests and larger files are handled as downloadable artifacts.