Weekly Trending

diagram-design is an agent skill for brand-aware editorial HTML/SVG diagrams, with website-style onboarding, source imports, and a fidelity ledger.

holaOS is an open-source, local-first agent workspace where Claude Code, Codex, and its own agent can share files, tools, apps, integrations, and editable persistent memory. It supports built-in or bring-your-own models and describes enterprise permissions and audit logs.

An open-source model-routing project that translates between incompatible provider interfaces and applies policy-driven routing, turning model choice into configurable infrastructure for cost, privacy, latency, and capability decisions.

Macro connects email, chat, documents, tasks, agents, calls, and CRM records through shared memory and a common data graph.

An open-source control plane for teams of AI agents, with goals, tasks, organisational roles, budgets, schedules, approval gates, and audit trails.

RL-Index moves part of retrieval reasoning from query time into the index, using LLM-generated rationales and reinforcement learning to improve retrieval quality and reduce latency.

A research paper on LLM agents choosing unnecessarily powerful tools and on evaluation and defences for least-privilege tool selection.

Many LLM agents keep their plans only in the active context window rather than as durable state. The paper argues that context management is a systems requirement for reliable long-running agents.

An empirical study finds that structured JSON-schema constraints can suppress tool calling in open-weight LLMs when grammar masks make tool-call tokens unreachable.

A matched comparison of GUI and CLI computer-use tasks finds that execution modality and skill coverage shape agent success as much as model capability: GUI agents fail through perception and grounding, while CLI agents fail when their skill layer is missing, brittle, or poorly routed.

This paper studies catastrophic failure in LLM multi-step tool use during reinforcement learning. It traces abruptly broken tool-call structures to probability spikes in control tokens and finds that interleaving supervised fine-tuning with reinforcement learning improves stability but remains brittle under format and content distribution shifts.