Automated AI News Brief: Agent Collaboration, Real-Time Voice, and Multimodal APIs
August 22 AI news brief: GitHub Copilot in Slack and Teams, the DeepSeek Vision API, low-latency Qwen3-TTS, and LLM and OpenRouter CLI updates.
Introduction
This brief was fetched by Horizon and selected and rewritten by Codex. Horizon only handled collection.
1. GitHub Copilot previews agentic work in Slack
GitHub brought agentic capabilities from Copilot CLI and the Copilot app into the GitHub integration for Slack, now in public preview. This lets teams work with @GitHub in an existing communication space, but they should first confirm channel permissions, repository access scope, and how outputs return to the normal review process.
Source: GitHub Changelog: The new GitHub Copilot experience in Slack
2. GitHub Copilot can also share agent sessions in Microsoft Teams
GitHub released shared agentic work in Microsoft Teams. Users can mention @GitHub in a channel, thread, or direct message to start a Copilot session that multiple people can view and direct. This can keep requirement clarification and agent execution in the same discussion context, but final modifications still need clear review ownership.
Source: GitHub Changelog: Shared agentic work with GitHub Copilot in Microsoft Teams
3. DeepSeek launches the v4 Flash Vision experimental API
DeepSeek published a guide for its v4 Flash Vision experimental release. Images are converted to tokens and billed alongside text tokens; before inference, images are automatically resized while preserving aspect ratio, with larger images reduced to roughly an 800×800 total pixel count. OCR and document-understanding products should test recognition quality and cost after this resizing with their own page layouts and text sizes.
Source: DeepSeek: v4 Flash Vision experimental guide
4. Qwen3-TTS implementation report targets sub-50ms time to first audio
Nari Labs shared latency optimizations for Qwen3-TTS, reporting 34ms p95 time to first audio at 10 requests per second on one H100, alongside the implementation, benchmark, and optimization notes. Real-time voice depends on more than audio quality: first-audio latency, concurrency, and hardware cost all matter. Treat this figure as a benchmark under specific conditions, not as a result every deployment can copy directly.
Source: Nari Labs: How we made a text-to-speech model respond in sub-50 ms
5. LLM and llm-openrouter updates address dependencies and tool calling
LLM 0.32.1 temporarily pins openai<3 to fix fresh installs that broke when a transitive httpx dependency disappeared; a later release is expected to move to httpx2. llm-openrouter 0.7 switches to OpenRouter's Responses API and adds three server-side tools: Shell, WebFetch, and WebSearch. These updates are a reminder that dependency changes and tool permissions both deserve to be treated as formal version-upgrade work in CLI agents.
Sources: Simon Willison: llm 0.32.1, llm-openrouter 0.7
Today's Notes
Today's trend is not simply making agents answer better, but bringing them into collaboration tools, voice interfaces, and multimodal workflows. The real quality bar remains whether permissions, cost, latency, and final review can be managed clearly.