Weekly Trending

An open-source context database gives agents a virtual filesystem for memories, resources, and skills, with tiered loading and observable retrieval paths.

A local desktop harness turns existing terminal coding CLIs into a coordinated multi-agent team with shared memory, mailboxes, worktrees, human approval gates, and an observable operator floor.

ai-memory is an open-source long-term shared-memory system for agent coding CLIs. It carries scoped task context across different agent harnesses instead of leaving it inside one chat interface.

Apache Maka is a local-first AI agent workspace that records model messages, tool calls, results, permission decisions, and termination events in an append-only operational log.

An Apache-2.0 Kubernetes control plane that suspends and resumes stateful agent sandboxes, multiplexing many agents across a smaller worker pool.

AI-Infra-Guard is an Apache-2.0 red-teaming suite for agent workflows, skills, MCP servers, AI infrastructure, and jailbreak evaluation, with CLI and CI-facing scanners.

CoffeeBench evaluates long-horizon LLM agents in a simulated coffee economy that requires communication, negotiation, pricing, inventory management, and transactions.

Qwen-Image-Agent proposes an image-generation workflow that builds missing context through planning, search, memory and feedback before rendering.

A study across 67 frontier models finds that routing, voting, cascades, and mixture-of-agents systems are bounded by the failures their constituent models share. Ensemble gains depend on genuinely different error patterns.

A paper that models LLM multi-agent systems as computational graphs and applies gradient-based connection weights for fine-grained credit assignment.

A paper on progressive multimodal search agents that choose among image search, text search, and stopping under explicit tool-call budgets.

A Korean benchmark that evaluates web agents on breadth search, exhaustive set enumeration, and table filling.

Google's paper proposes an AI-assisted scientific-review tool and a framework for levels of AI-human collaboration in review.

AgentOdyssey procedurally generates long-horizon text-game environments for agents that learn during test time through exploration, episodic memory, and planning.