Harness Evolution · Trend Harvest
KeystoneHub가 외부 AI agent 생태계에서 본 신호를 수집 → 분석 → 거절/적용으로
나눠 항목별로 기록합니다. 적용은 rule·skill·hook·docs 중 하나로 실제 반영된 항목입니다.
스냅샷 기준 2026-05-26 (run16)
Ground self-configuration in a capability graph of the agent's OWN primitives + deterministic floor + sandbox-promote-with-revalidation, so the agent cannot propose config using primitives that do not exist.
Only verified (execution/measurement-sourced) values may enter an artifact; unverified numbers are sanitized at an output gate.
AI 출력 품질은 코드 리뷰로 못 잡는다 — 출력 자체를 회귀 테스트 케이스로 고정해야 측정·차단 가능
Symbolic short-term memory: condense heavy tool logs into compact Mermaid symbols (−61% tokens, +51% pass rate); layered long-term memory distills conversations into personas/scenes instead of flat vectors.
Prefix-cache stability: keep the prompt prefix byte-identical across turns so provider KV/prompt cache hits stay high, keeping token cost flat over long sessions.
Make any software agent-native via community CLI registry with JSON+Human dual output and verify tests; agents call CLIs instead of learning per-app APIs.
Secure persistent personal agent server in Rust: one binary, sandboxed execution, 15 lifecycle hooks, session branching.
Cursor adopts the same manifest+skill+MCP primitive triad as Anthropic/microsoft-apm, confirming convergence on a portable plugin spec across AI coding agents.
Aggregator repo with no extracted patterns at this moment; revisit when TOC reorganization completes.
Lightweight constantly-evolving personal AI agent.
Product CLI for Google Workspace with bundled agent skills; not a reusable harness principle.
Open-source Rust agent harness, multi-provider, runs locally; 'sovereign by design'.
Agent harness performance optimization system. Skills + instincts + memory + security + research-first dev for Claude/Codex/Opencode/Cursor.
Production-grade gateway managing unified access to GenAI services on Envoy - declarative routing/auth/rate-limit for multi-provider LLM.
GEP-powered self-evolving engine — auditable evolution with Genes, Capsules, Events. Agent evolution as auditable record.
Self-evolving agent that grows skill tree from a 3.3K-line seed achieving full system control with 6x less token consumption.
Git scraper pattern - state changes captured as git commits over time, zero-infra change-tracking.
Git-scraping primitive applied to social/sync domain.
Same git-scraping primitive applied to FEMA disaster data.
Git-scraping for Florida outages.
Git-scraping GitHub Actions package versions - track dependency drift via git diff.
Git-scraping for USGS data.
ARIS — Lightweight Markdown-only skills for autonomous ML research with cross-model review loops and idea generation, paper review, paper writing automation.
Self-configuring agent that *ingests its own runtime + docs* to build a capability graph. Agent introspection -> auto-skill-discovery.
Static analysis for GitHub Actions security - HARD primitive for CI safety.
Production-grade engineering skills for AI coding agents (Addy Osmani / Google).
Claude Code GitHub Action — CI integration: PR comments, scheduled runs, self-healing test loops.
AST-based structural code search/lint/rewrite CLI. Tree-sitter-driven exit-code-friendly.
'Making ALL software agent-native' - universal CLI wrapping platform.
Pre-indexed code knowledge graph -> caller/callee graph + symbol DB cuts repeat grep across same codebase by 94%.
12 factors for production agents: own prompts, own context, own control flow, compact errors, stateless reducer steps, etc.
Skill that reduces token usage by 65%. HARD-measurable token efficiency.
AI agents running training research autonomously on single-GPU - keep/discard loop in production at 82K stars.
Codex variant of autoresearch — modify/verify/retain/discard loop.
Clean LLM output of 'slop' (verbose AI-style filler). Token-efficiency primitive.
Claude Autoresearch Skill — autonomous goal-directed iteration inspired by Karpathy.
Hash-anchored edits + LSP integration + deterministic ast_edit are concrete primitives that make agent code-edits a HARD-checkable operation: stale anchor → reject before file corruption, instead of letting the LLM retry. Same family as our semantic-entity-diff.md but extends to the WRITE side.
Terminal coding agent with hash-anchored edits — content hash verification of file region before applying mutation. Primitive for safe agent code mutation.
Claude Code Dedicated Development Harness achieving high quality through autonomous cycles — name aligns directly with our harness concept.
Claude/Codex skill for deliberate skill development during AI-assisted coding — detects "learning opportunities" in real coding flow and steers the user toward them.
Hash-anchored edits = HARD primitive for safe agent code mutation. LSP support + subagents = production harness.
Persistent context across sessions captured as hooks. Production-grade session-context capture.
Stealth Chromium that passes bot detection - drop-in Playwright replacement.
Frontend stack for agent-native UI with React and Angular — agent-native UI components.
Meta-prompting + context engineering + spec-driven dev. Wave execution + discuss mode + gate taxonomy.
LangGraph variant of autoresearch loop - structured graph-based research orchestration.
Multi-stage research pipeline (research -> write -> review -> revise -> finalize). FSM-based research workflow.
LLM Council - multi-model anonymous review + Chairman synthesis.
Tony Pocock's .claude skills directly published; ranks #2 weekly Shell. Skills-as-portable-units pattern continues to gain.
Agent specialization with 'reality-checking agent' role - blind validation layer in production.
AI SKILL providing design intelligence for professional UI/UX across multiple platforms.
Local-first open-source Claude Design alternative with 19 Skills and design system generation.
Agentic skills framework + software-dev methodology with two-stage review gate (Spec Compliance + Code Quality).
'Bash is all you need' - minimal agent harness from 0 to 1. Educational/reference.
Open source agent that lives on machines 24/7 keeping apps running.
RAG -> autonomous reasoning agent -> self-maintaining Wiki. Wiki auto-maintenance is novel angle on memory hygiene.
Build autonomous AI agents with reliability primitives + MCP + universal commerce protocol.
Agent capability graph - what the agent can do, encoded as graph.
Long-memory eval suite for ActiveGraph. Eval is HARD by definition (numeric scores).
Super AI assistant with task planning + skill creation.
Anthropic-published CLI for Claude API.
Claude Agent SDK for Python.
Claude Agent SDK for TypeScript.
Anthropic-managed plugin directory. Official curation source.
Plugins for knowledge workers (Jupyter, docs, etc).
Anthropic public skills repository — canonical reference for SKILL.md format.
Enterprise platform: guardrails + MCP registry + gateway.
Claude/Codex/Gemini API proxy. Unified gateway in Go.
Agent harness framework.