Research Brief

The Agency (github.com/msitarzewski/agency-agents) is an open-source collection of 147 AI agent personalities across 12 divisions, born from a Reddit thread and months of iteration. Installation is straightforward: copy agents to your Claude Code directory with cp -r agency-agents/* ~/.claude/agents/. The project includes an installer script that scans your system for installed tools (Claude Code, Copilot, Antigravity, Gemini CLI, OpenCode, Cursor, Aider, Windsurf, Qwen Code) and lets you select which to install via a checkbox UI. Each agent is specialized with unique voice, communication style, and deliverable-focused workflows.

Promptfoo is an AI security platform that OpenAI announced acquiring on March 9, 2026, with plans to integrate it into OpenAI Frontier. The platform is trusted by over 25% of Fortune 500 companies and includes a widely-used open-source CLI and library for evaluating and red-teaming LLM applications. The GitHub repo (github.com/promptfoo/promptfoo) allows you to test prompts and agents, perform red teaming/pentesting/vulnerability scanning for AI, compare performance of GPT, Claude, Gemini, Llama, and more with simple declarative configs and CI/CD integration. The testing framework allows teams to define prompts, expected outputs, and evaluation criteria in configuration files that can be automatically run against one or more AI models. Founded in 2024, Promptfoo began as an open-source framework for evaluating AI prompts and model behavior, later expanding into a commercial platform used by developers and enterprise security teams. Prior to acquisition, Promptfoo had raised $23.6 million in funding, including an $18.4 million Series A round in July 2025.

MiroFish (github.com/666ghj/MiroFish) is described as "A Simple and Universal Swarm Intelligence Engine, Predicting Anything". It's a next-generation AI prediction engine powered by multi-agent technology that extracts seed information from the real world (breaking news, policy drafts, financial signals) and automatically constructs a high-fidelity parallel digital world where thousands of intelligent agents with independent personalities, long-term memory, and behavioral logic freely interact and undergo social evolution. Setup requires filling in API keys in .env file for LLM API configuration (any OpenAI-compatible API), LLM_BASE_URL, LLM_MODEL_NAME, and ZEP_API_KEY for Zep Cloud configuration. MiroFish has received strategic support and incubation from Shanda Group, and its simulation engine is powered by OASIS. The README recommends starting with fewer than 40 rounds to manage LLM API costs; version 0.1.0 was released in December 2025; the project topped GitHub's Global Trending list in March 2026 and attracted investment from Shanda Group founder Chen Tianqiao. The OASIS research paper notes that LLM agents tend to be more susceptible to herd behavior than real humans, and simulated crowds can polarize faster than real ones.

Impeccable (github.com/pbakaus/impeccable) is described as "The vocabulary you didn't know you needed: 1 skill, 20 commands, and curated anti-patterns for impeccable frontend design" with a quick start at impeccable.style to download ready-to-use bundles. Anthropic created frontend-design, a skill that guides Claude toward better UI design; Impeccable builds on that foundation with deeper expertise and more control. It includes an expanded skill with 7 domain-specific reference files covering typography, color-and-contrast, spatial-design, motion-design, interaction-design, responsive-design, and ux-writing, plus 20 steering commands to audit, review, polish, distill, animate, and more, along with curated anti-patterns that explicitly tell the AI what NOT to do. Installation is simple: npx skills add pbakaus/impeccable auto-detects your AI harness and places files correctly. For project-specific installation: cp -r dist/claude-code/.claude your-project/ or global installation: cp -r dist/claude-code/.claude/* ~/.claude/. It works with Cursor, Claude Code, Gemini CLI, and Codex CLI.

OpenViking (github.com/volcengine/OpenViking) is an open-source context database designed specifically for AI Agents (such as openclaw) that unifies the management of context (memory, resources, and skills) through a file system paradigm, enabling hierarchical context delivery and self-evolving. OpenViking is initiated and maintained by ByteDance's Volcano Engine Viking team. OpenViking has a built-in memory self-iteration loop where developers can actively trigger the memory extraction mechanism at the end of each session; the system will asynchronously analyze task execution results and user feedback and automatically update them to the User and Agent memory directories, extracting operational tips and tool usage experience from task execution, allowing the Agent to get "smarter with use" through interactions with the world. It features tiered context loading: L0 (abstract), L1 (overview), L2 (details) to reduce token usage and load content on demand. The project has 13.3k+ GitHub stars and is actively maintained with commits within the last 35 minutes as of recent reports. It was open-sourced in January 2026.

Heretic (github.com/p-e-w/heretic) is a tool that removes censorship (aka "safety alignment") from transformer-based language models without expensive post-training, combining an advanced implementation of directional ablation (also known as "abliteration") with a TPE-based parameter optimizer powered by Optuna, enabling it to work completely automatically. Heretic finds high-quality abliteration parameters by co-minimizing the number of refusals and the KL divergence from the original model, resulting in a decensored model that retains as much of the original model's intelligence as possible. Installation requires Python 3.10+ with PyTorch 2.2+ installed: pip install -U heretic-llm then run heretic Qwen/Qwen3-4B-Instruct-2507 (replace with whatever model you want to decensor). On an RTX 3090 with default configuration, decensoring Llama-3.1-8B-Instruct takes about 45 minutes. Heretic supports model quantization with bitsandbytes to drastically reduce VRAM required; set the quantization option to bnb_4bit to enable quantization. Over 1,000 Heretic models have been created and published by the community on Hugging Face.

Nanochat (github.com/karpathy/nanochat) is the simplest experimental harness for training LLMs, designed to run on a single GPU node with minimal/hackable code covering all major LLM stages including tokenization, pretraining, finetuning, evaluation, inference, and a chat UI. You can train your own GPT-2 capability LLM (which cost ~$43,000 to train in 2019) for only $48 (~2 hours of 8XH100 GPU node), or on a spot instance for closer to ~$15. The entire pipeline is contained in the single file runs/speedrun.sh, designed to be run on an 8XH100 GPU node, taking ~3 hours to run. The goal is to improve the state of the art in micro models accessible on budgets of < $1000; it's not an exhaustively configurable LLM "framework" but a single, cohesive, minimal, readable, hackable, maximally-forkable "strong baseline" codebase designed to run start to end and produce a ChatGPT model you can talk to. The official repository is hosted on GitHub under Andrej Karpathy's account, and the entire codebase is open source and freely available for research, education, and commercial use. It's approximately 8K lines of clean, dependency-light code built for learning, hackability, and rapid modification by researchers.