Cobus Greyling Claim this profile

Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.

Practical reference for fleet engineering, governing populations of AI agents with accountability

The canonical reference for Grok Build CLI /goal — run-until-done objectives, patterns, verifier skills, and goal-audit CLI.

Controllable reasoning demos for NVIDIA Nemotron 3 Super (120B/12B MoE) — chat UI, CLI, API server, tool calling, budget sweep, and adaptive routing

A compounding knowledge base maintained by LLM agents — inspired by Karpathy's LLM Wiki pattern

Personal collection of Agent Skills for AI coding agents (Claude Code, Gemini CLI, Cursor, Codex, etc.)

LlamaIndex Notebooks

Claude Code Agent Teams — Multi-agent orchestration from your terminal (Opus 4.6)

Long-term memory layer for AI agents

NVIDIA Nemotron 3 Nano Omni — blog, meeting intelligence agent, and Gradio UI

OpenAI-compatible proxy that routes requests to the best NVIDIA NIM model by task

Production-grade skills, custom MCP servers, and real projects built with Grok Build 0.1 (grok-build-0.1). The definitive toolkit + living showcase for xAI's agentic coding model. Sole author & maintainer: Cobus Greyling

Minimal Agent Harness — Six harness components demonstrated with NVIDIA Nemotron 3 Super

Own the Outer Loop. Evidence → Verdict → Answerability. Governance layer for agentic engineering.

Agent = Model + Harness — Markdown conversion of the Substack article by Cobus Greyling, including all original images

Amazing CLI utilities for Grok Build by Cobus Greyling. 7 power tools: sessions, skills, backup, usage, mcp, worktree, memory.

Building Specialised AI Agents with the Claude Agent SDK — blog and 8 demo agents

4 viral-grade experiments built on xAI Grok: brutal voice roasts, canonical realtime voice client, voice-driven image gen, and the best X thread generator

Runnable illustrations of Assistant (on-behalf-of) vs Claw (fixed-identity) agent authorization

Factorial benchmark for agentic search: retriever × delivery × harness (inspired by Is Grep All You Need?)

My Universal Agent Thesis — Full digital autonomy requires agents that build their own tools, discover their own boundaries and operate any system they encounter.

Composable harness runtime for production agents — primitives, traces, evolution, outerloop governance.

NVIDIA Nemotron-Terminal — The CLI Is The Path To AI Autonomy

Three approaches to building AI Agents with Claude — SDK, Markdown, and Teams

A collection of MCP servers.

HuggingFace notebooks

Markdown memory hygiene for agent wikis, trust decay, context packing, and archive gc

What Andrej Karpathy learnt — a distilled field guide

What is NVIDIA SpatialClaw? Simple explainer blog + conceptual demo of code as the action interface for spatial reasoning.

Real-time gaze safety monitor demo for human-in-the-loop agent oversight

NVIDIA Nemotron 3 Ultra showcase — interactive demos and benchmarks

Grok Build CLI - article and screenshots

NVIDIA Riva voice agent — ASR + LLM + TTS with Gradio web UI

NeMo AgentIQ — Enterprise Agentic RAG with Multi-Source Reasoning. Powered by NVIDIA NIM & LangChain/LangGraph.

opus 4.8 utilities

Blog: The CLI Is The Path To AI Autonomy — connecting NVIDIA terminal study, Steinberger's CLI-over-MCP position, and Anthropic's Vercept acquisition

LangGraph multi-agent workflow examples (fork; archived)

Runtime safety layer for NVIDIA NIM-powered agents — composable guardrails for input validation, PII detection, tool-call auditing, and cost controls

A secure, always-on runtime for autonomous agents — gateway, policy guardrails, OpenShell sandbox, and durable skills. Dependency-free reference implementation.

10 progressive, production-minded tutorials for learning to engineer reliable AI agent systems and harnesses with LangChain. Sole author: Cobus Greyling.

Reasoning-backed guardrails for Nemotron agent deployments

Right-Sizing AI Agents: Why Specialization Beats Scale — Blog, code examples, and architecture diagrams exploring NVIDIA Nemotron 3 specialized model stack for production agentic AI

Web companion dashboard for Claude Code Agent View — live status, telemetry, and per-session post-mortem

A minimal, fully-local agent harness for Microsoft Fara-7B — the 7B-parameter agentic small language model for computer use.

A living, executable showcase and practical toolkit for Grok Build (xAI). Includes reusable skills, exemplary AGENTS.md, a Grok-readiness auditor CLI, and demonstrations of Plan Mode + subagents.

Practical patterns, starters & CLI tools for memory engineering with AI coding agents. Stop re-explaining the repo — design the memory. Get a Memory Ready score.

Instruction Fade-Out — Demonstrating how LLM instruction compliance degrades over long conversations and how event-driven reminders fix it

Comparative document analysis with transparent chain-of-thought using NVIDIA Nemotron Super 49B

The Token is Becoming the New Hidden Compute Primitive — Tokens are following the same path as CPU clock cycles, from headline metric to invisible infrastructure

Gradio app for browsing, chatting with, and comparing NVIDIA NIM models

LangGraph research partner with visible, reviewable, git-backed memory

Replace MCP with CLI — why the command line may be the best AI agent interface

Self-improving agent loop demonstrating emergent harness components — HyperAgents meets Harness Engineering

Markdown version of the blog post: Is The Missing Piece in AI Agent Tools?

Intent Laundering — test whether your AI safety guardrails are real or cosmetic. Based on Golchin & Wetter (2026).

Claude Code Hooks: Deterministic Control Over Non-Deterministic AI — blog, demo code, and diagrams

Blog: Claude Code Agent View — one screen for every background session

AI Agents Are Better at Building From Scratch With Less Context — Minimal prompts outperform detailed prompts for code generation

The AI Agent Reality Gap Just Got Wider — article by Cobus Greyling

AI Agent Benchmark Suite — 5 real-world coding tasks to benchmark any AI coding agent. One suite. Every agent. Real scores.

AI Agent Safety & Governance Auditor — Intent Laundering, Instruction Fade-Out, Policy Testing, OWASP Agentic Top 10, Compliance Reports

The open source coding agent.

LlamaIndex is the leading document agent and OCR platform

A list of AI autonomous agents

OpenAI API notebooks and early experiments (legacy; archived)

Example Bot Showing the weather with basic filfullment

Visible Reasoning Agent Framework for NVIDIA Nemotron. Every model thought, incremental tool call, and decision is logged and surfaced in real time.

Pi_k,m,r — propose/critique/compare orchestration that lets a weak reasoning model match frontier performance on verifier-backed tasks.

AI-Powered Workflow Automation - Describe workflows in plain English, FlowPilot wires up the APIs automatically.

LLM-in-Sandbox with Agent Observation & Monitoring Dashboard

Four ways to build AI agents with Claude — Agent SDK, Markdown Definitions, Agent Teams, and Managed Agents — same task, four paradigms

Classify and analyze AI agent session failures across frameworks

Zero-dependency Python graph runtime for agent loops with Mermaid export

No description.

Skeleton-of-Thought Prompting — outline first, expand in parallel. Blog and demo code.

AI Agents & The Skill Boundary Problem — exploration-first execution and skill boundary awareness based on the OSExpert paper

Blog: PSTN Is The New CLI — voice as the universal agent interface

Blog post: an LLM agent whose memory is a folder of markdown files, reviewable and git-backed

Blog: Two-thirds of the harness is not the model — where agent gains actually live

Build LLMs from scratch — educational notebooks and exercises

The Singularity Is Dead, Intelligence Grows Like a City - Blog post on why AI intelligence grows socially, not individually

LangChain sandbox demos — native Pyodide+Deno isolation and Sandbox-as-Tool pattern for AI agents

LLM Self Improvement

The Evolution of Shared Language in AI Agent Development - From prompts to harness engineering

Generate, score, filter, and export synthetic training data using NVIDIA Nemotron models. Over 98% of Nemotron's alignment data is synthetic — this implements that pipeline.

Why 2.6 million AI agents could not form a culture

Building with Claude's Extended Thinking — Blog and demo code

Basis chatbot no NVIDIA GPU

Remote MCP server without authentication — experimental prototype

Where Does OpenClaw AI Agents Actually Fail? A trajectory-based safety audit blog post.

The Mediation Layer: Why Language Models Are Becoming the Universal Interface

Blog & interactive demo: NVIDIA OpenShell infrastructure-level security for autonomous AI agents

GPT-5.5 Computer Use Agents — OpenAI released GPT-5.5 April 2026 with native computer use. The model gives you vision. The harness gives you agency.

From Copilot to Codex — The Shift from Autocomplete to Delegation

NVIDIA Nemotron Nano 2 VL for invoice understanding, internal audit and fraud detection

Synthesis of 17 arXiv papers examining the Moltbook AI-agent-only social network (February 2026)

Model Distillation Attacks — how distillation is being weaponised to steal AI IP

The Real Moat in AI Is Distribution — why critical mass beats technology every time

Architecting Agentic AI — How SDKs, Scaffolding, Frameworks & Harnesses differ, and why harness engineering is its own discipline

Why AI Agent Teams Work Best With Minimal Structure — based on Dochkina (2026) 25,000-task study on multi-agent coordination

The agent harness is where coding-agent gains actually live — notes on Agentic Harness Engineering

GPT-5.4 Native Computer Use — The first general-purpose model that truly operates your computer

98% of Claude Code Is Not AI - Blog analyzing a 46-page study reverse-engineering Claude Code's architecture

The Four Debts of Agentic AI — blog post and interactive demo based on Vishnyakova (2026)

Blog: When The IDE Becomes Optional — The Development Environment Is The Next Layer To Collapse

Configured, not coded. The engineering discipline gap in agent development.

Companion repo to my Medium article on AI Agent Traps — the Google DeepMind taxonomy of six attack surfaces against AI agents.

Death of the Demo: How autonomous agents are reshaping the sales cycle

Eat Your Own AI — The Dogfooding Test

Context Engineering Is the Real Product — blog and interactive demo based on the OpenDev technical report

The AI Agent Reality Gap — Analysis of the gap between demo and deploy for LLM-based agents in production

Agentic Governance with Bounded Autonomy — Blog, Gradio demo app, and images based on OpenAI's Agentic Governance Cookbook

Sandboxing — How Codex uses sandboxes across the Codex app, IDE, and CLI

Claude Sonnet computer-use demo — minimal agent harness for browser/desktop control

Mirror of the Substack post Agents of Chaos by Cobus Greyling

Integrate cutting-edge LLM technology quickly and easily into your apps

Visualise, explore, and calibrate your RAG pipelines

Created for toolchain: https://console.bluemix.net/devops/toolchains/cc18819e-cf7f-4143-8c6f-de2d4788f1d1?env_id=ibm%3Ayp%3Aeu-de

Learn how to use CUA (our Computer Using Agent) via the API on multiple computer environments.

NVIDIA plugin for secure installation of OpenClaw

Official inference framework for 1-bit LLMs

A curated list of awesome skills, hooks, slash-commands, agent orchestrators, applications, and plugins for Claude Code by Anthropic

AI Agent Governance Toolkit — Policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous AI agents. Covers 10/10 OWASP Agentic Top 10.

Remote MCP server without auth — superseded by remote-mcp-server-authless-1 (archived)

DSPy Notebooks

SmartLLMChain is a LangChain implementation of the self-critique chain principle. It is useful for particularly complex question answering and follows a cycle of ideation, critique and resolve.

LangChain Projects

Haystack Notebooks

This is a very basic Amazon Lex bot illustrating integration with a Lambda function.

Automatic Prompt Engineer (APE)

Cohere Notebooks

Early LLM API examples and snippets (legacy; archived)

LLM drift detection experiments (legacy; archived)

IBM Speech-to-Text language model experiments (legacy; archived)

Rasa banking chatbot demo (legacy; archived)

LangChain cookbook recipes and runnable examples

Convert PDFs and images to LLM-ready Markdown using NVIDIA Nemotron 3 Nano Omni

Runtime orchestration framework for NVIDIA Nemotron 3 Nano Omni — adaptive compaction, doom-loop detection, system reminders, modality-aware inference

A markdown-first cookbook showcasing Claude Opus 4.7 features: /effort, extended thinking, parallel tool use, memory, subagents, and skills.

LangChain integration for the Kore.ai XO Platform — retriever, tools and document loader