Weekly Trending

Brainhuggers Bureau
← PrevKW31 · 27.07–02.08.2026Next →
GitHub Trending
01
huggingface/speech-to-speech
+177 this wk · 7.2k stars
huggingface/speech-to-speech: modular local voice-agent stack
A low-latency voice-agent pipeline that separates VAD, speech-to-text, language model, and text-to-speech components while presenting an OpenAI Realtime-compatible WebSocket API and allowing local model options.
02
ColeMurray/background-agents
+9 this wk · 2.2k stars
ColeMurray/background-agents
An open-source background coding-agent system for work that continues after the interactive session closes.
03
modelcontextprotocol/ext-apps
MCP Apps / An Interface Layer for Tool Calls
The MCP Apps protocol and SDK let MCP servers deliver interactive interfaces such as charts, forms, or canvases inside compatible chat clients. A tool response can carry a UI resource that the host renders in a sandboxed iframe.
04
safishamsi/graphify
77k stars
Graphify / A Queryable Knowledge Graph for Agent Context
Graphify is an AI coding assistant skill that converts folders of code, schemas, scripts, documents, papers, images, and videos into a queryable knowledge graph.
HF Papers
01
Train Harness-native Agents in Any Environment
Train Harness-native Agents in Any Environment
OpenForge RL trains agents inside the real inference harnesses they will inhabit, including coding tools, browsers, GUI agents, and computer-use environments.
02
JarvisHub
JarvisHub: canvas-native multimodal creative agents
JarvisHub is an open harness for canvas-native multimodal creative agents, using an editable canvas to hold artifacts, dependencies, versions, feedback, and tool actions as shared state for people and agents.
03
arXiv:2606.22936
When Agents Commit Too Soon / Diagnosing Premature Commitment in LLM Agents
Researchers study premature commitment in long-horizon LLM agents: an agent settles on an early interpretation and then defends it. They use cross-run hidden-state convergence as a diagnostic for trajectory consistency; it does not tell whether the agent is correct.
04
Capable but Careless
Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity?
A paper evaluating whether computer-use agents respect contextual boundaries when working across tools such as email, calendars, and task systems.