Weekly Trending

Atlas links coding-agent commits to the sessions, prompts, tool calls, reasoning, file changes, and shared local memory that produced them.

SkillRise uses a single policy to solve tasks while curating an evolving, transferable skill file that carries reusable operating knowledge from one task to the next.

A paper evaluating whether GUI agents can learn procedural tasks from video tutorials, using task-aware keyframe extraction to improve performance.

A multi-agent retriever, verifier, critic, and generator loop improves text-rich image data by using rejected samples as feedback.

HExA learns reusable skills through active experimentation instead of relying only on fixed retrieval, parametric knowledge, or one-shot search.

AutoSaddler uses traces from agent runs to propose narrow harness updates and retains them only after validation. It treats the agent's surrounding control logic as a repairable system rather than a fixed wrapper.

Recuris separates an agent's live working state from a repairable archive of experience and learned skills. The architecture aims to let long-running agents accumulate useful experience without treating every old transcript as active context.

Constraint Weakening studies how binding requirements can lose force as they pass through LLM-agent workflow handoffs. A rule can remain present in a summary while becoming optional in the next action.

The AFTER benchmark evaluates whether procedural memory in LLM agents transfers across workplace tasks, professional roles, and model backbones.

A survey of agentic artifact creation that frames dependable outputs as operational representations with construction policies and runtime verification.

A paper on constraining self-evolving LLM agent harnesses so changes remain recoverable and earlier states can be reconstructed.