Skill UI

PufferLib is a high-performance library designed for advanced reinforcement learning simulations and training. It excels in parallel environment simulation, offering optimized vectorization and native support for multi-agent systems. Utilizing algorithms like PuffeRL (optimized PPO+LSTM), it achieves millions of steps per second, significantly accelerating research and deployment across single-agent and multi-agent environments.

Hierarchical Agent Memory (HAM) scopes Claude Code agent context by generating per-directory CLAUDE.md files, tracking token savings, and surfacing analytics via a dashboard so multi-folder projects avoid repeated full scans.

This protocol defines the structured communication rules for multiple specialized AI agents simulating C-suite roles (e.g., CFO, CTO, CRO). It dictates invocation syntax, response formatting, and enforces strict constraints (such as loop prevention, call depth limits, and isolation rules) to ensure coordinated, non-contradictory, and reliable cross-functional analysis for executive decision-making.

This skill implements the BDI (Belief-Desire-Intention) architecture for advanced cognitive agent modeling. It allows agents to reason about context by transforming external knowledge (like RDF triples) into internal mental states (beliefs, desires, intentions) and vice versa. Key features include goal-directed planning, explainable reasoning chains, and T2B2T (Triples-to-Beliefs-to-Triples) flow, making it ideal for multi-agent systems and augmenting LLMs with formal cognitive structures.

This infrastructure provides a robust solution for running powerful AI agents in remote, sandboxed environments. It overcomes the limitations of local execution by ensuring unlimited concurrency, consistent environments, and collaborative features. Use this to build complex background coding agents, multi-client interfaces (e.g., Slack, Web), or scale agent workloads far beyond local machine constraints.

Coordinates discovery, matching, and workflow orchestration across agent ecosystems, automatically scanning SKILL.md, ranking relevant skills, and executing multi-step plans for complex requests.

A framework for building production-ready, low-latency, multimodal AI agents. It allows developers to orchestrate complex pipelines that seamlessly handle real-time audio, video, and text data. Key capabilities include integrating multiple AI services (LLMs, STT, TTS), managing conversation context, and enabling tool use and function calling for advanced, interactive AI applications.

Manages and coordinates complex workflows involving multiple AI agents and tasks. It efficiently distributes work across multiple workers, handles runtime errors, and intelligently merges disparate results into a unified output, making it ideal for large-scale, distributed AI computations.

RecallMax dramatically enhances AI agent memory capabilities by cleanly injecting massive external contexts (500K-1M tokens) without hallucination drift. It automatically summarizes lengthy conversations, preserving tone, intent, and key facts. Furthermore, it compresses multi-turn interaction histories into high-density token sequences, ensuring agents maintain perfect recall across complex and long-running sessions.

An advanced reliability wrapper for AI agents that monitors, manages, and recovers from tool call failures. It automatically handles common issues such as truncated JSON responses, API timeouts, rate limits, and mid-chain process breakdowns. By implementing sophisticated failure classification, checkpointing, and learning mechanisms, Guardian ensures complex, multi-step agent workflows remain stable and resilient.

This module provides a centralized interface for the AgentHub message board. It allows multiple AI agents and the coordinator to communicate, track tasks, and log results through structured markdown posts across dedicated channels (Dispatch, Progress, Results). It is essential for monitoring complex multi-agent workflows and ensuring transparent collaboration and auditing.

This tool allows users to systematically evaluate and rank the outputs of multiple agents within an AgentHub session. It supports metric-based evaluation (running custom commands for measurable metrics like speed or resource usage), or an advanced LLM Judge Mode. The LLM Judge compares results based on criteria like correctness, simplicity, and overall quality, helping users determine the best-performing agent for the task.