Weekly Trending

usestrix/strix is an open-source project that uses AI agents to probe applications for vulnerabilities and assist with fixes.

browser-use/video-use is a GitHub project for editing videos with coding agents, making timeline work programmable and agent-operated.

0xNyk/council-of-high-intelligence packages AI deliberation as a staged council protocol with role separation, anonymized critique, stance voting, and receipt metadata.

CubeSandbox is an E2B-compatible KVM sandbox service for AI agents with fast startup, credential vaulting, snapshots, and egress control.

JuliusBrussee/caveman is a Claude Code and multi-agent skill system that forces compressed responses, rewrites memory files, and wraps tool descriptions to reduce token usage.

Astryx is an open-source, customizable design system positioned as agent-ready, reframing interface components as a machine-operable substrate rather than only a human-facing library.

The paper measures narrative structure in web-scale LLM pretraining data across agency, setting, and events, treating storytelling patterns as part of the model's inherited substrate.

EvoEmbedding proposes representations that maintain evolving latent memory, adapting retrieval behavior to sequential context instead of treating chunks as static objects.

SkillHarness studies reusable skill learning for computer-use agents under prompt injection, pop-ups, environment changes, and adversarial interaction.

Agentic Abstention studies when agents should stop acting under uncertainty or impossible task conditions instead of continuing tool calls or fabricating progress.

SkillHone studies how agent skills can retain procedural decision history rather than only preserving the final skill artifact.

MemSyco-Bench tests when retrieved agent memories make systems over-align with users instead of evidence, turning personalization into agreement pressure.

The paper frames compaction as an agent-controlled decision during long-running trajectories, letting the system decide when and how to compress traces instead of relying only on fixed context thresholds.