MiroFish is a multi-agent prediction engine. Its core idea is to take real-world “seed materials” — such as news, policy drafts, financial signals, reports, or even stories — and build a parallel digital world where thousands of agents with memory, personality, and behavioral rules interact to simulate future scenarios. The system is designed to output both a detailed forecast report and an interactive environment for exploring the results. It works in five main stages: graph construction / GraphRAG, environment and character setup, parallel simulation, report generation, and deep interaction with agents and the ReportAgent. So it is not just a chatbot giving opinions; it is framed as an agent-based social simulation system. In practice, it looks like a full application rather than just a paper or prototype. The repository includes backend, frontend, Docker files, an .env.example, and run scripts. It appears to launch both frontend and backend together, with the frontend on port 3000 and the backend API on port 5001. It also requires Node 18+, Python 3.11–3.12, uv, LLM API keys, and Zep Cloud. Other public projects from the same author include: BettaFish — a multi-agent public opinion analysis / media monitoring system, positioned as the upstream data and analysis layer.MiroFish — the simulation and prediction engine.MindSpider — an AI crawler for public opinion analysis, currently archived.DeepSearchAgent-Demo — a demo for building a Deep Search Agent from scratch.StateMachineDrawing — an online tool for drawing finite automata.LateralMovement_DetectionSystem — an older project focused on lateral movement detection.666ghj/666ghj — the author’s profile repository. The most notable point is that MiroFish seems like the natural continuation of BettaFish. BettaFish focuses more on data collection and public opinion analysis, while MiroFish moves up a level into general simulation and forecasting. Together, they appear to form a pipeline from raw data to predictive decision support.

End of MCP? A new tool called CLI Anything may change how AI agents interact with software. It allows you to transform any open-source software into a command line interface (CLI) tool that can be controlled directly by AI agents. Instead of using programs by clicking through graphical interfaces, the tool automatically analyzes the software’s source code and generates terminal commands to control it. This makes it possible for AI agents to perform tasks in those programs without needing APIs, complex integrations, or fragile automation setups. The process is simple: 1. Install CLI Anything 2. Point it to the software repository 3. The tool analyzes the code 4. It automatically generates the CLI, tests, and documentation 5. In practice, this means AI agents could control applications such as graphic editors, audio software, development tools, or any other open-source program directly from the terminal. This opens the door to much more powerful automation and agents that can truly operate real software. Repository link:https://github.com/HKUDS/CLI-Anything

# Multi-Agent AI Architecture — Summary The document describes a **multi-agent AI architecture** where different AI models handle different types of tasks, coordinated by a **central orchestrator**. The main idea is: > Use the **best model for each task** instead of relying on a single model. This improves **cost efficiency, speed, and capability**. --- # System Flow Messages are received through a **Telegram bot**, then processed by an **Ollama orchestrator**, which decides how to handle the request. Basic flow: ``` User → Telegram Bot → Ollama Orchestrator → Selected Agent → Response ``` Users can also explicitly select an agent using commands: ``` /claude /codex /ollama /openrouter ``` If **no command is provided**, Ollama automatically decides which agent to use. --- # Role of the Ollama Orchestrator Ollama acts as both: **1. Router (Orchestrator)** It analyzes the message and decides whether to: * respond directly * delegate the task to another agent **2. Local AI Agent** It can also answer simple requests itself. This reduces costs because **local models are free**, while paid models are used only when necessary. --- # Task Classification The orchestrator classifies requests into three types: ### Simple tasks Examples: * questions * conversations * translations → handled directly by **Ollama** --- ### Code or tool tasks Examples: * editing files * running commands * debugging → routed to **Codex or Claude** --- ### Complex tasks Examples: * large refactoring * deployment * complex analysis → routed to **Claude** --- # Agents in the System ### Claude Agent Command: ``` /claude [model] ``` Models: * Opus * Sonnet * Haiku Capabilities: * bash execution * file editing * web search * advanced tools Limitation: usage quota. --- ### Codex CLI Command: ``` /codex ``` Used mainly for **automated coding tasks**. Runs as: ``` codex --full-auto "message" ``` Capabilities: * code generation * file editing * command execution

https://github.com/inematds/bs-benchmark https://inematds.github.io/bs-benchmark/viewer/index.v2.html

ClawSec: Security Framework for OpenClaw 🔒 🏴‍☠️ If you're still deep in the OpenClaw rabbit hole and concerned about security, there’s a new option worth knowing about. Prompt Security has released ClawSec — a framework with five skills that you can provide to your agent so it can protect itself. Here’s the breakdown: - Embedded Advisory for monitoring malicious skills - Soul Guardian for policy enforcement - Watchdog for full configuration monitoring - Incident reporting - Real-time advisory feed This matters because many people are pulling skills from random repositories, and some of them come with built-in malware. You can use these tools as a foundation to make your OpenClaw instances increasingly secure over time. https://github.com/prompt-security/clawsec