Agentic CEO – 스스로 탐색하고, 비판하고, 진화하는 AI 연구 유기체

An autonomous multi-agent research system that acquires knowledge, builds a persistent worldview, and improves itself. 3,700+ knowledge entries. 173 research hunts. 7 specialized agents. 35 days of autonomous operation. ~$24/month. A system of specialized AI agents that work together to research, extract, organize, challenge, and synthesize knowledge — autonomously. It ran daily from March 3 to April 6, 2026, accumulating a structured knowledge base across 68 domains without human intervention beyond setting initial goals. Not a chatbot. Not a RAG pipeline. A research organism. Read the full technical deep-dive: How I Built a Multi-Agent Research System That Ran Autonomously for 35 Days → ┌─────────────┐ │ DIRECTOR │ Plans, prioritizes, resolves │ (the CEO) │ └──────┬──────┘ │ ┌──────▼──────┐ │ KNOWLEDGE │ 3,700+ entries │ BASE │ 68 domains │ (blackboard)│ All agents read/write └──┬──┬──┬──┬┘ │ │ │ │ ┌──────────┘ │ │ └──────────┐ │ │ │ │ ┌─────▼───┐ ┌─────▼──▼──┐ ┌──────▼────┐ │ WOLF │ │ JACKAL │ │ NEXUS │ │ (hunts) │ │(scavenges)│ │(connects) │ └─────────┘ └───────────┘ └───────────┘ │ │ │ │ ┌─────▼─────┐ │ │ │ CRITIC │ │ │ │(challenges)│ │ │ └───────────┘ │ │ │ ┌─────▼───────────────────────────▼─────┐ │ INWARD CYCLE │ │ Observer → Critic → Evolver │ │ (watches) (challenges) (improves) │ └───────────────────────────────────────┘ All agents communicate through a shared file-based knowledge base — not by talking to each other. This means: - Observable: Every thought is a file. You can read every decision the system made. - Loosely coupled: Agents don't know about each other. Add or remove agents without rewiring. - Persistent: Nothing is lost. The knowledge base is version-controlled. - Auditable: Trace any conclusion back to its source, extraction, and critique. | Agent | What It Does | Key Detail | |---|---|---| | Director | Sets daily research priorities based on knowledge gaps | Reads the constitution, reviews what's missing, writes missions | | Wolf | Hunts for answers using a 5-phase predator protocol | Stalks → test-bites → commits or abandons → extracts → synthesizes | | Jackal | Scavenges Wolf's kills for overlooked insights | Never searches the web — re-reads what Wolf found with fresh eyes | | Nexus | Finds cross-domain connections across the knowledge base | "3 independent evidence lines point to the same opportunity" | | Dissolve | Strips complexity theater from regulations and standards | Turns 200-page FSMA documents into actionable guides | | Equalizer | Democratizes insider knowledge | Makes expert-level information accessible to non-experts | | Bridge | Closes the gap between knowing and doing | Produces step-by-step action plans from research findings | | Agent | What It Does | Key Detail | |---|---|---| | Observer | Measures everything — kill rates, cost per insight, source reliability | Grades each hunt A through F | | Critic | Challenges assumptions, flags blind spots, demands evidence | Reviews any framework rated above 0.9 confidence | | Evolver | Implements improvements based on Critic's recommendations | Adjusts search strategies, tunes thresholds, reallocates budget | The core research engine. Not a web scraper — a predator. Phase 0: TERRITORY (5% budget) What do we already know? What gaps remain? Phase 1: STALK (15% budget) Broad search. Score results against known gaps. Select 5-7 targets. Ignore everything else. Phase 2: TEST BITE (20% budget) Fetch first 1,000 characters of each target. Commit or abandon. Most sources are thin. Walk away fast. Phase 3: KILL (50% budget) Full extraction on committed targets only. Value hierarchy: Organs (contradictions) > Meat (gap-fillers) > Bones (confirmations) Phase 4: FEED (10% budget) What changed? Which gaps closed? Which opened? Where should the next hunt go? Real metrics from 173 hunts: - Average kill rate: 40-60% (targets committed / targets stalked) - Average cost per hunt: $0.13–$0.30 - Average frameworks per hunt: 10-20 - Best source: arxiv.org (96.6% success rate, 5.67 frameworks/fetch) - Dead sources identified: medium.com, reddit.com (0% extraction across all attempts) The system doesn't just research what it's told to. It maintains a list of things it finds genuinely interesting: - Complex adaptive systems and emergence - History of failed predictions by experts - Biomimicry in engineering - Mathematical paradoxes and what they reveal about logic - The history of railroad standardization (and what it teaches about protocol design) - How blind cavefish adapt — and what it says about vestigial systems in software Twice a week, the Director picks a curiosity topic, runs a full research cycle, and the Essayist writes a long-form essay. The results are published at brcrusoe72.github.io/directors-notes. Sample hunt: How do blind cavefish repurpose visual neural tissue? — 4 kills, 18 frameworks, 8 organs, $0.18. Overturned the "repurposing" narrative: the tectum retains its excitatory architecture while selectively losing inhibitory circuits. The pleiotropic package hypothesis is dead. The system has values. They're enforced, not decorative. intellectual_honesty: - Never suppress contradicting evidence - Confidence scores must reflect actual uncertainty - "I don't know" is a valid and valuable output anti_echo_chamber: - Actively seek opposing viewpoints on any topic with >5 frameworks - Critic must review any framework rated >0.9 confidence - Flag when all sources on a topic share the same bias autonomy_boundaries: - May acquire and analyze content freely - Must NOT act on conclusions without human approval - Must NOT spend money without pre-authorized budgets | Metric | Count | |---|---| | Knowledge base entries | 3,704 | | Research hunts completed | 173 | | Domains covered | 68 | | Agent reports generated | 61 | | Pipeline summaries | 20 | | Nexus executive briefs | 22 | | Observer daily reports | 7 | | Published essays | 4+ | | Total cost | ~$25-30 | technology · strategy · manufacturing · ai-systems · finance · psychology · military-training · trust-mechanisms · procurement · supply-chain · food-safety · restaurant-operations · philosophy · economics · agriculture · nutrition · statistics · regulatory · contrarian - Python 3.12+ - An Anthropic API key (Claude) - AgentSearch running on localhost:3939 (for web search) git clone https://github.com/brcrusoe72/agentic-ceo.git cd agentic-ceo pip install -r requirements.txt export ANTHROPIC_API_KEY=your_key_here python tools/hunter.py "What are the actual deployment costs of UNS/OPC-UA at 100-machine scale?" python tools/orchestrator.py This runs: Director → Wolf hunts → Jackal scavenging → Nexus synthesis → Barrier cycle → Observer → Critic → Evolver python tools/director.py # Plan today's priorities python tools/hunter.py "your question" # Single research hunt python tools/jackal.py --last-hunt # Scavenge the latest hunt python tools/nexus.py # Cross-domain synthesis python tools/observer.py # Performance measurement python tools/critic.py # Challenge assumptions python tools/curiosity.py # Pick something interesting to learn python tools/essayist.py # Write an essay from a curiosity hunt - ARCHITECTURE.md — Original system design and rationale - ARCHITECTURE-V2.md — The closed loop: outward + inward + immune cycles - WOLF_JACKAL.md — Wolf & Jackal predator architecture (replaces linear Hunter) - DEEP_DIVE.md — Full technical blog post: "How I Built a Multi-Agent Research System" - Cavefish Neural Reallocation Hunt — A curiosity hunt that overturned a hypothesis - Executive Brief: April 6 — What the system knew after 35 days - Observer Daily Report — Self-measurement: hunt grades, source reliability, kill chain stats - Curiosity Interests — What the system finds interesting - Blackboard > message passing for observable multi-agent systems. Files are debuggable. Conversations are not. - The Wolf protocol's test-bite phase saves ~40% of extraction budget. Most web content is thin. Test before committing. - Jackal (lateral scavenging) finds things Wolf can't. Focused research has blind spots. A second pass without a specific question consistently surfaces overlooked connections. - The Critic is the most important agent. Without adversarial review, knowledge bases become echo chambers. Requiring review of high-confidence frameworks prevents premature certainty. - Curiosity-driven research produces disproportionately interesting outputs. The cavefish hunt and the railroad standardization essay were both curiosity-driven. They connected to manufacturing and protocol design in ways no directed research would have found. - Cost scales with knowledge, not compute. At $0.80/day, the bottleneck is never the API bill — it's whether the system is asking the right questions. - Claude (Opus for strategy/critique, Sonnet for extraction/synthesis) - AgentSearch (self-hosted search API) - OpenClaw (agent orchestration) - Hugo (essay publishing) - Python, JSON files, and stubbornness MIT — the architecture and agent code are open. The knowledge base contents (3,700+ entries) are not included in this repo. Built by Brian Crusoe · Crusoe Advisory