Weekly Trending

trycua/cua is open-source infrastructure for computer-use agents, including desktop sandboxes, SDKs, and benchmarks across macOS, Linux, and Windows.

OpenMontage is an open-source agentic video production system with pipelines, media tools, and hundreds of skills for automated content creation and editing.

BuilderIO's agent-native framing makes the interface shift explicit: applications are becoming structured habitats for autonomous operators.

Graph Memory for LLM Agents proposes reconstructing long-horizon agent memory through associative graph exploration rather than static retrieve-then-reason lookup.

HarnessX proposes a composable, adaptive, and evolvable agent harness where prompts, tools, memory, and control flow are typed primitives assembled and improved from execution traces.

The paper frames the shift from chatbots to persistent autonomous AI systems with workspaces, skills, verification loops, governance, and state-action-observation trajectories.

RedAct studies how agent execution traces can be redacted to protect procedural skills while preserving useful auditability and evaluation evidence.

Data2Story frames data journalism as a multi-agent production system with analysis, evidence inspection, narrative angle, and multimodal presentation as routed roles.

VisualClaw combines streaming perception, retrieved memories, hot/cold skill injection, and skill evolution for a personalized physical-world agent.

WebStep evaluates web agents through semantic trajectories, skill decomposition, and bifurcation analysis instead of only terminal success.

PhoneHarness routes phone-use agents across GUI, CLI, and structured tools while checking observable side effects.

Selective Control under Noisy Perception shows that aggregate governance metrics can look healthy while harms concentrate on users bridging communities.

Dr-DCI turns retrieval into dynamic corpus interaction: agents pull documents into an expandable workspace, then search, filter, compare, and verify inside it.