Weekly Trending

ArtCraft is a visual IDE for AI image and video creation that combines 2D and 3D scene composition with multiple generative-model providers. It makes camera, asset placement, scene layout, and model choice controllable before rendering.

A GitHub project for natural-language end-to-end web and mobile testing: an agent completes a goal, assertions check the result, and a successful action trace can replay without further model calls.

An open-source company AI workspace that combines agents, sandboxed user-built applications, and Gatekeepers for scoped permissions, audit logs, and deferred human approval.

This paper examines filesystem-based memory for LLM agents, including how files are organized, changed, and maintained over time. It highlights that retrieval quality degrades when the store is left without explicit stewardship.

A multimodal agentic reasoning approach that keeps a structured evidence ledger and constrains claims to tool-produced records, making provenance part of the reasoning process rather than an after-the-fact citation step.

This paper examines how credible-looking but false knowledge can enter deep-research workflows and lead to false conclusions. It argues that isolated verification can miss contamination that persists through a long research trajectory.

A framework for agents to model beliefs, intentions, norms, and partial knowledge when planning communication.

A method for evaluating whether an edited image remains consistent with multiple visual references.

An agent harness that externalises task state and verifies each step against the environment for long-running tasks.

Zero-Mem proposes performing memory operations by retrieving raw traces with entity and temporal structure rather than using model tokens to rewrite a memory narrative. The paper focuses on preserving original evidence while reserving model calls for answering.

A paper proposing an interaction-centric taxonomy for locating failures in agent systems, distinguishing problems in the model, harness, tools, user interaction, and environment.

A benchmark for evaluating LLM agents over long-term e-commerce operations, where prices, supplier choices, cash flow, and delayed outcomes compound across decisions.

A benchmark for recursive self-improvement in personal agents across experience, memory, routines, and learned skills, requiring evidence that prior work produces better later work.