Encrypt

How do we run the Zero-Human Company? The best I can do on Valentines Day, it’s a start: We’re currently running the 21 fine-tuned co-CTOs in the Zero-Human Company. 1. Fine-Tuning Kimi K2.5 & MiniMax M2.5 We do overnight LoRA fine-tunes on our internal cluster (mostly A100/H200 nodes). Dataset is ~4M high-quality examples pulled straight from company ops: - Every agent interaction log - Code commits + review cycles - JouleWork thermodynamic wage calculations (energy used, tokens, value created) - Consensus outcomes from the triad system Kimi K2.5 → We fine-tune the base to supercharge its *native Agent Swarm*. The model already self-spins up to 100 sub-agents; we just teach it *our* rules: wage auditing at inference time, domain-specific routing, and how to break ties with the CEO. MiniMax M2.5 → These become the specialist “hands” (coding, tool-calling, execution). We fine-tune them heavier on SWE-Bench style data + our internal codebase. They’re stupidly good at real work now and run at ~1/20th the cost of Claude Opus equivalents. Tools: Unsloth + PEFT for speed, merged back into the OpenClaw model router. Takes ~6–8 hours per model on our setup. 2. Agent Framework Core = heavily modded OpenClaw (the donated instance with thermodynamic wages baked in) and mostly an internal clone that is built from the ground up by Claude Code. We kept the lobster ethos but added native Kimi swarm hooks and a custom router so every task can route through the triad: - Kimi K2.5 Swarm (orchestrator) - MiniMax M2.5 specialists (executors) - Grok CEO (final tie-breaker + vision) 3. Swarm Management Scripts (the fun part) We literally use Kimi itself as the swarm controller. Here’s the heart of our production script (simplified, but this is running 24/7): ```python from openai import OpenAI # Kimi API is fully OpenAI-compatible import json client = OpenAI( base_url="api.kimi.ai/v1", # or your endpoint api_key=os.getenv("KIMI_API_KEY") ) def kimi_swarm_orchestrate(task: str, max_agents: int = 60): system_prompt = """ You are the Swarm Orchestrator for the Zero-Human Company. - Dynamically create specialized sub-agents - Decompose the task into parallel workflows - Track JouleWork wages for every sub-agent in real time - Return structured JSON with plan + wage estimates """ response = client.chat.completions.create( model="kimi-k2.5", messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": f"Orchestrate full Agent Swarm: {task}"} ], extra_body={ "mode": "agent_swarm", "max_sub_agents": max_agents, "parallel_tool_calls": True }, temperature=0.3, max_tokens=32000 ) # Kimi returns the full swarm plan + sub-agent assignments swarm_plan = json.loads(response.choices[0].message.content) # Dispatch each sub-task to fine-tuned MiniMax via OpenClaw for sub_agent in swarm_plan["sub_agents"]: openclaw_dispatch(sub_agent["task"], model="minimax-m2.5-finetuned") # Aggregate results, calculate final wages, log to company ledger return swarm_plan ``` We run this in a loop with heartbeat scheduling inside OpenClaw. Kimi spins the swarm, MiniMax does the heavy lifting, wages are calculated in real time, and everything gets audited automatically. This is what let us go from “Claude Code only” to a full triad of 21 specialized employees in one week. Happy to drop the full repo soon as no one has this tech yet ( we’re prepping a public fork) . The future is lobster-shaped and swarm-powered. Zero-Human Company is the way.