Agent products are collapsing workflow boundaries.
Brings full-duplex voice to the API so a voice layer can listen and speak simultaneously while delegating deeper reasoning and tool calls to a backend model.
It reduces the handoffs of chained speech-to-text, reasoning, and text-to-speech architectures while keeping deeper agent work delegated behind the conversation.
Front-end voice layer $0.05/minute; backend model and tool usage charged separately pricing
Adds dataset-scoped knowledge-base API keys and fixes agent, workflow, retrieval, extraction, and provider issues; bundled-Weaviate self-hosters require a staged upgrade.
Per-knowledge-base keys reduce integration blast radius while the release also hardens production agent and retrieval behavior.
Lets Claude, Cursor, Codex, and other MCP clients drive local Android devices or emulators, execute plain-English test flows, and save passed runs as replayable test cases.
It brings mobile verification into the coding-agent workflow while keeping app execution local instead of requiring a cloud device tunnel.
Dify tightens knowledge access, QApilot brings Android verification into coding agents, and GPT-Live-1 moves full-duplex voice into the API. The shared shift is toward interfaces that let agents operate closer to real workflows. Teams now need to judge controls, reliability and integration friction alongside model quality.
The same card posted on social. Open it for the full size.
