Weekly Trending

ByteDance's deer-flow repository presents a SuperAgent harness for research, coding, and creation with sandboxes, memories, tools, skills, subagents, and message gateway surfaces.

A local git proxy validates AI-generated code in a disposable worktree before forwarding changes and opening a clean pull request.

Page-agent explores natural-language control over web interfaces, treating pages as surfaces agents can inspect and operate.

MemSlides proposes a slide-generation agent that separates user-profile memory, working memory, and tool knowledge so presentations can be revised locally across multiple turns.

GateMem benchmarks shared-memory agents on recall, authorization boundaries, active forgetting, and multi-principal governance.

StylisticBias studies how small controlled visual style cues can shift social judgments made by multimodal language models.

OpenRath proposes session-centered runtime state for agents, making transcripts, tools, memory, replay, branching, and audit state into an inspectable object.

Language world models simulate agentic environments before the agent acts, using large interaction trajectories across multiple domains.

The paper frames agent memory as a data-management system with representation, extraction, retrieval/routing, and maintenance modules rather than a loose transcript store.

The paper argues that verifying coding-agent work is becoming harder than generating it because scalable rewards remain proxies for underspecified human intent.