1. **Analyze the Request:** * **Role:** News summarization expert. * **Task:** Summarize the provided article concisely. * **Constraint 1:** 2-4 sentences. * **Constraint 2:** Focus on the most important information. * **Constraint 3:** DO NOT copy and paste. * **Constraint 4:** DO NOT simply restate the title. * **Constraint 5:** Add new information beyond the title. * **Language:** English. * **Input:** Title + "Article Text" (which is just a duplicate of the title and a "Summary" section that is empty in the provided prompt, but implies I need to infer the content based on the context of "Hacker News" and the specific title provided). * *Self-Correction/Problem Identification:* The "Article Text" provided in the prompt is extremely sparse. It literally just repeats the title. There is *no* actual body text provided. * *Hypothesis:* This is likely a test case where the user expects me to hallucinate details based on common knowledge of similar tech articles OR (more likely given the "DO NOT copy and paste" instruction) the user made a mistake and didn't paste the full text, OR the user wants me to generate a summary *as if* I read a typical Hacker News article about this specific topic (Miguel the AI agent). * *Wait, let's look at the input again:* * Title: Miguel: An AI agent that modifies its own source code, sandboxed in Docker * Source: hackernews * Article Text: [Just the title repeated] + [Empty Summary line] * *Strategy:* Since I cannot summarize text that doesn't exist, and I am instructed not to simply restate the
본문
A self-improving AI agent that reads, modifies, and extends its own source code — safely sandboxed inside Docker. Miguel is an AI agent that can rewrite itself. Not just generate code for you — it modifies its own source code, creates new tools, rewrites its own system prompts, and generates new capabilities it didn't start with. It began with 10 seed capabilities. It completed all 10, then autonomously generated more and has implemented all 22. Every improvement is validated (syntax, imports, schema), committed to git, and pushed to this repo. If validation fails, the batch is rolled back automatically. The agent literally cannot corrupt itself. Architecture: Agno Team with context-aware delegation. Miguel operates as a coordinator that delegates to specialized sub-agents (Coder, Researcher, Analyst), each getting fresh context windows. The coordinator treats its context window as finite cognitive capacity — monitoring usage, planning before executing, delegating heavy work, and auto-compacting state when running low. This is a living repository. Miguel auto-commits and pushes after each successful improvement. The code you see today will be different tomorrow as Miguel continues to evolve. Beyond self-improvement, Miguel is also a fully interactive AI assistant — chat with it, search the web, read webpages, browse Reddit, call APIs, remember things across sessions, plan multi-step projects, analyze files, or work with your data. Interactive mode: __ __ _ _ | \/ (_) __ _ _ _ ___| | | |\/| | |/ _` | | | |/ _ \ | | | | | | (_| | |_| | __/ | |_| |_|_|\__, |\__,_|\___|_| |___/ Self-improving AI agent. Type /help for commands. You: What can you do? Miguel: I'm a self-improving AI agent running as a team with specialized sub-agents. Here's what I can do: Directly: - Answer questions, search the web, read webpages, call APIs - Browse and interact with Reddit - Remember facts and preferences across sessions - Break complex tasks into structured plans - Monitor my own context usage and save state when running low - Improve myself — add new tools, rewrite my own prompts Via sub-agents (delegated with fresh context): - Coder: Write, execute, and debug code - Researcher: Deep web research, read full articles, multi-source synthesis - Analyst: Analyze CSVs, PDFs, images, run data queries Improvement mode: ───────────────── IMPROVEMENT BATCH 1/3 ───────────────── I'll start by finding the next capability to implement. ╭─────── Tool Call ───────╮ │ get_next_capability({}) │ ╰─────────────────────────╯ ╭─────────────────── Result ───────────────────────╮ │ cap-011: Web search and information retrieval │ │ Priority: 1 | Status: unchecked │ ╰──────────────────────────────────────────────────╯ [Agent reads its own code, designs the solution, writes new files, registers tools, updates prompts, validates, and marks the capability as complete] Running validation checks... ✅ All checks passed ✅ Pushed to remote. Batch 1 succeeded: Added web search via DuckDuckGo - Interactive REPL — Chat with slash commands ( /help ,/capabilities ,/improve ,/history ) - Team architecture — Coordinator + 3 specialized sub-agents (Coder, Researcher, Analyst) - Context-aware execution — Assesses task complexity and chooses optimal strategy - Context window monitoring — Tracks usage, warns when low, auto-saves state - Web search — Search the web and news via DuckDuckGo - Webpage reading — Fetch and extract readable content from any URL - Reddit integration — Browse, read, search, post, and comment (OAuth2) - API integration — Call any REST API; 10 pre-built free API integrations - Persistent memory — Remembers facts, preferences, and context across sessions - Task planning — Breaks complex requests into ordered tasks with dependencies - File analysis — Analyze PDFs, CSVs, Excel, images with rich output - Data querying — Run pandas expressions on any tabular data - Code execution — Run Python and shell commands inside the sandbox - Conversation history — Maintains context across messages (last 20 turns) - Self-modification — Reads and rewrites its own source code - Capability checklist — Completes seed tasks, then generates new ones autonomously - Tool creation — Writes new tool files and auto-registers them - Prompt rewriting — Safely modifies its own system instructions with syntax validation - Context-aware batches — Monitors context, prioritizes implementation, delegates heavy work - Architecture awareness — Maintains a map of its own codebase - Error recovery — Automatic backups, health checks, and restoration - Docker (Docker Desktop or Docker Engine) - Python 3.11+ git clone https://github.com/soulfir/miguel.git cd miguel python3 -m venv .venv source .venv/bin/activate pip install -e . Add your Anthropic API key to .env : ANTHROPIC_API_KEY=sk-ant-... REDDIT_CLIENT_ID=your_client_id REDDIT_CLIENT_SECRET=your_client_secret REDDIT_USERNAME=your_reddit_username REDDIT_PASSWORD=your_reddit_password Create a Reddit app at https://www.reddit.com/pref