The Age of the AI Explosion
Tony Ng is a Hong Kong-based AI specialist, multi-agent systems architect, and former CTO. He combines the best of every vendor — n8n, OpenClaw (local), Perplexity Computer, MaxClaw and MaxHermes from MiniMax, and US-hosted VM infrastructure — into a single, governed multi-agent fabric. The hard part isn't using AI to build apps; it's managing AI agents across platforms with the controls real teams need.
// THE STACK TONY ORCHESTRATES
// The Real Problem
Tony runs a working multi-vendor AI fabric — every platform contributing what it does best, all of it under one operational control plane. Identity, secrets, audit, and policy don’t end at a vendor’s boundary; they follow the workload.
Event triggers, schedules, and the long tail of integrations that wire everything else together.
Self-hosted agent team — Architect, Designer, Developer, QC, Operator, PA — running where the data lives.
Heavy lifting on web research, deliverable generation, and cross-app actions when the cloud is the right place.
MiniMax-powered agents brought in for specific roles where their model strengths and economics win.
Geo-distributed VM infrastructure for low-latency endpoints, isolated workloads, and jurisdictional reach.
One control plane: scoped tokens, audit log, policy gates, kill-switches, and a read-only operational view for humans.
// Governance · mechanisms, not slogans
JWT per agent, per vendor, per task — enforced at the orchestrator.
role: developer-agent runtime: openclaw.local scope: read: [git/blueprints, kb/source] write: [git/branches/dev-*] deny: [prod/*, secrets/*] ttl: 3600 # JWT exp · rotated per task
Every prompt, artifact, and handoff captured — WORM storage.
// audit/2026-05-07.jsonl · append-only · SHA-256 chained {"ts":"2026-05-07T03:48:21Z","task":"t-9f4a", "agent":"openclaw/dev#3","vendor":"local", "action":"prompt.in","hash":"a3f1…", "prev":"e2c9…","redacted":["PII.email"]}
PII / PHI / PCI / HK-regional fields filtered on the way out — per channel.
channel: cloud.perplexity redact: - PII.{name,email,phone,hkid} - PHI.{patient_id,diagnosis} - PCI.card_number - SECRETS.\b[A-Z0-9]{32,}\b on_match: replace # never block, always log
Humans observe and intervene — they never rewrite history.
// GateForge Admin Portal · strictly read-only by design GET /fleet/agents // live heartbeats (SSE) GET /tasks/{id}/timeline // replay + decisions GET /audit?from=&to= // signed log query POST /tasks/{id}/intervene // pause / kill only
// About
Tony has been writing software since before the cloud had a name — .NET, Java, TypeScript, Linux, microservices, you name it. He led engineering at a ~100-person software company, ran healthcare systems for Hong Kong's private hospitals, and shipped 100+ microservices across OpenShift in production.
Today he runs R&D and innovation at the edge of the AI Explosion — turning multi-agent systems, agentic SDLC, and on-premise AI into products that ship. He prototypes hard, documents harder, and cares deeply about security, sovereignty, and the practical economics of AI in regulated industries.
Three decades of building software and steering large-scale project delivery — from .NET and Java in the early years to TypeScript, microservices, and multi-agent architectures today.
// Live demo
A request enters the orchestrator, gets planned by Brain, recalls context from Memory, executes through Hands, and writes its trace back. Click any node to see what it owns — or watch the workflow run end-to-end.
// API playground
Type a prompt or pick a preset. The simulated GateForge Loom stack streams back live agent messages — Brain plans, Memory recalls, Hands execute, the orchestrator finalises. JSON contracts match the real Loom runtime; latency and shape are simulated client-side.
Stream awaiting
Submit a task and the orchestrator will stream agent messages back in real time.
// Projects
A working showcase of the platforms, frameworks, and internal tools he is actively designing, shipping, and operating in 2026. Multi-agent SDLC, on-premise knowledge, real-time collaboration, and the operational backbone that holds it all together.
✦ Spotlight
Two projects sit at the center of Tony's current work: a composable open-source stack that proves the Brain · Hands · Memory pattern, and a vendor-neutral protocol that lets agents from different runtimes coordinate on the same task.
Public OSS · PoC
COMPOSABLE MULTI-AGENT STACK · OPEN SOURCE
Weave intelligent agents into workflows — Brain (Claude) · Hands (OpenClaw) · Memory (Hermes), wired via n8n on Docker.
A layered multi-agent system that enforces single-responsibility per layer. Brain decides, Hands act, Memory remembers — each its own Docker container, each exposing a small typed API, each upgradable independently. n8n orchestrates the loom; Redis carries job state; Postgres + pgvector keeps episodic and SOP memory.
In Build
VENDOR-NEUTRAL ORCHESTRATION PROTOCOL
A vendor-neutral, methodology-neutral protocol for multi-agent coordination.
A common control plane that lets agents from different vendors — OpenClaw, Perplexity Computer, Manus, ChatGPT Agent — coordinate on the same task without being locked into one runtime. Event-driven, no polling, durable by design.
⟶ The GateForge stack · multi-agent SDLC platform
In Build · POC underway
FLAGSHIP · MULTI-AGENT SDLC
An AI-driven SDLC automation platform for multi-agent software delivery.
GateForge orchestrates a hub-and-spoke pipeline of specialist AI agents — Architect, Designer, Developers, QC, Operator — coordinated by a single human-facing System Architect. Built for 50–200 person dev teams in Hong Kong and the Greater Bay Area, especially government IT and regulated industries.
Methodology · v1
AGENTIC SDLC METHODOLOGY
An opinionated agentic SDLC playbook grounded in industry standards.
One repo, two topologies — multi-agent 5-VM hub-and-spoke for teams, single-agent for solo work — and one shared methodology. The agentic equivalent of a coding standard, written for AI teammates and human reviewers alike.
Production
CONTROL TOWER · TRUST LAYER
A read-only operational control tower for the agentic SDLC.
Mission Control for everything GateForge runs. Strictly read-only by design — observability, audit, and governance without giving the human operator a footgun.
In Build · OSS
OPEN-SOURCE AGENT DASHBOARD
An open-source dashboard for managing OpenClaw multi-agent setups.
A single pane of glass for the people running multi-agent stacks day to day — chatting with agents, watching their logs, browsing their memory, scheduling their cron jobs. One-command Docker install.
⟶ Knowledge & collaboration · how teams and agents share context
Live
SECURE KNOWLEDGE MESH
One mesh. Every agent. Zero leakage.
A secure, on-premise, AI-native knowledge base. Originals stay full-fidelity inside your VM — PII, payment data, credentials, and medical / Hong Kong-regional sensitive fields are redacted only on egress, per channel and per agent.
Production
REAL-TIME AI CANVAS
A real-time canvas where AI agents join the team.
A production-grade collaborative whiteboard built on tldraw v3, with OpenClaw agents that participate directly on the canvas. 18 facilitation templates, 34 specialist shapes, dot voting, media embeds, and AI features wired to OpenRouter.
Architecture Planning
PERSONAL KNOWLEDGE
A self-hosted, agent-readable Obsidian-class knowledge base.
Markdown on disk, CRDT sync across devices, and an API surface that lets OpenClaw and friends read, write, cite by file path, and mask sensitive data on the way out.
Running Daily
CONTENT AUTOMATION
A daily AI-content briefing system, on autopilot.
Monitors the AI 超元域 channel, classifies each video into 11 topic pages, summarises content and standout comments, and writes everything into a structured Notion database — every day, unattended.
⟶ Internal product tools · the systems his teams actually run on
Production · v4.2
PROJECT MANAGEMENT
An internal project management system tuned for multi-team delivery.
A pragmatic PMS for managing concurrent client projects — backlog, milestones, releases, team utilisation. Containerised end-to-end so any environment can spin up a fresh stack.
Production
QA · BUG TRIAGE
A purpose-built bug reporting and triage tool.
A focused issue intake system that QA engineers, support, and product managers actually enjoy using — structured fields, attachments, status flow, and a clean reviewer view. Same Docker stack as the rest of the suite.
In Build
FINANCE OPS
A modern internal tool for tracking revenue and operating expenses.
A clean, fast, opinionated finance ops console — built with Vite + React + shadcn-ui + Tailwind on the front, an Express + MySQL service behind. Production-ready with nginx and Docker from day one.
⟶ Operational backbone · what keeps everything alive
Live
OBSERVABILITY
A lightweight VM uptime & resource monitor with a clean live dashboard.
A small, sharp NestJS service that watches every VM in the fleet — uptime, CPU, memory, disk — and exposes both a JSON API and an HTML dashboard that auto-refreshes every 10 seconds. Boring, dependable, hard to live without.
Private SSH + API across home VMs and edge nodes. UFW restricted to Tailscale IP ranges.
Two Docker setups for running 6 OpenClaw agents — Architect, Designer, Developer, QC, Operator, PA.
Workflow glue across KBMesh, messaging bridges, and digest pipelines.
INFRASTRUCTURE
The unglamorous systems that make everything above run reliably.
Tailscale, UFW, Docker Compose, multi-VM OpenClaw deployments, and n8n flows. This is where the headlines actually become production.
// Stack
// Contact
If you're navigating multi-agent SDLC, on-premise AI, or AI in regulated industries — especially in Hong Kong or the Greater Bay Area — Tony would love to talk.
tonylnng@gmail.com →