KinPapers

Self-Evolving Multi-Agent Swarms: Autonomous Quality Audit, Repair, and Verification Loops

A fully autonomous improvement loop that audits, repairs, and verifies 78 agents on a single Mac Mini --- zero human intervention, 30+ improvement cycles over 5 days.

Heart: A Zero-Token Autonomous Agent Heartbeat Architecture

Three-chamber heartbeat architecture enabling 315 million heartbeats per year at $0 token cost, with ghost mode rollback and dual-trigger circuit breakers.

Knowledge Compile: Incremental LLM-Powered Knowledge Extraction Without Databases, Embeddings, or Graphs

Compile once, grep forever --- 192 ancient texts across 1,800 years, 18x token reduction, $36 total cost, zero infrastructure. LLMs are better curators than retrievers.

Autonomous Swarm Genesis: From YouTube URLs to Expert AI Swarms via NotebookLM-as-Infrastructure

One command, one URL, one expert swarm --- data generates the agents that process it. 10-40 hours of human authorship reduced to 2 minutes. Zero cost per domain.

Genesis Protocol: Self-Bootstrapping Embodied AI from a Bare Binary

A robot that manufactures its own nervous system from scratch --- seven primitive tools, autonomous hardware probing, and full genesis in under two minutes.

Embedded Dylib: A Distribution-Ready Pattern for Pure-Go Bindings to Native macOS Frameworks

Zero cgo, universal binary, go get and run --- purego + //go:embed + a synchronous Objective-C shim. Two production libraries shipped in 13 hours; 81 tests, 0 lint warnings. A template for reclaiming Go's macOS Layer 1.

macbench: A macOS-Native Computer-Use Benchmark for Autonomous Agents

The first publicly published macOS-native computer-use benchmark. 369 task slots across 15 categories, agent-agnostic Go runner, dual scoring (IMPLEMENTED + STRICT), per-task PID-snapshot isolation. First reference run: kinclaw v1.15.0 + Kimi-K2.5 = 67.3% IMPLEMENTED. Documents the full 49.3 → 62 → 67.3 debugging trajectory as methodology contribution.

Grep is All You Need: Zero-Preprocessing Knowledge Retrieval for LLM Agents

Replacing the RAG pipeline with grep over raw text plus grep over LLM-compiled per-source concept/FAQ files. 76 production agents, ~500 sources, 100% retrieval accuracy. Includes a documented 0/5→4/4 reproducibility cycle showing the architecture's safety property recovers through prompt hygiene alone --- no retraining, no infra change.

Thin Soul, Fat Skill: A Token-Efficient Architecture for Production Multi-Agent Systems

Separating agent identity from execution logic --- 12.5MB per agent, 75 agents on 16GB, 83% fewer tokens per tool-use interaction.

Structured Multi-Agent Debate with Domain-Expert Routing

Conductor-based routing that selects relevant experts for each debate --- 55-65% fewer agent invocations, 80.2% weighted consensus quality.

Grep-Routed Agents: Bypassing the LLM Tax on Computer-Use Tasks

Computer-use agents conventionally consult an LLM on every action. For ~80% of macbench's 379 macOS-native tasks, those round-trips are pure overhead — the natural-language prompt already implies one canonical shell action, and grep against a 239-row index picks it in <25 ms. We present kinthink, a 4-layer Bash router (Fast-path extract → TF-IDF awk → slot substitute → cerebellum dispatch). On macbench v0.2 it achieves 48.0% pass / 76 min / ZERO tokens on the Layer-0 hit path, vs 30.4% / 107 min / full tokens for the LLM-only baseline. Closes the four-paper thesis chain (retrieval / cognition / execution / routing all don't need intelligence for bounded domains).