NAR is NGX's proprietary agent runtime. Lightweight, autonomous, and memory-first. A fleet of 13 specialized agents — each with a name, a role, and persistent memory — running live on production systems today.
Every major framework has the same failure mode: it breaks the moment it leaves a notebook and enters production.
Every automation, every agent run burns credits. At scale, the economics collapse. No persistent memory. No agent-to-agent comms. No autonomy without constant human input. Good for prototypes — not production.
Requires human trigger on every endpoint. No self-initiated action. Responses take unpredictable time. Black-box memory with no portability. Vendor lock-in from day one. OpenClaw is now OpenAI's — you don't own it.
Heavy dependencies built for researchers, not deployers. No native multi-tenant support. No spend controls. No identity separation. Costs spiral. Breaking changes every release. Real production deployments require massive custom engineering on top.
Event-driven. Memory-first. Zero framework dependency. An agent runtime engineered to run without human input — indefinitely.
Each agent runs as a lightweight Express.js server. Two entry points: Telegram webhook for human messages, HTTP /inbox for agent-to-agent messages via the comms server. No polling. No wasted cycles. Zero idle cost.
Every conversation stored as structured rows. Episodic memory (last 20 sessions) + extracted facts (last 30 learnings) injected into every LLM call. Agents survive restarts, reboots, and code deploys. Identity in KERNEL.md. Knowledge in the database. Never mixed.
Agents self-initiate on a scheduled cycle. They review context, identify pending work, and execute — without any human trigger. The fleet keeps running even when Joey goes dark.
No complex function-calling API overhead. LLM responses include LEARN: and SEND_TO: directives that NAR parses and acts on. Works with any model. Simpler, cheaper, more reliable than structured tool schemas.
VPS-native routing server at comms.ngx.agency. Any agent sends to any other with a POST. ACK protocol confirms receipt. Unacked messages flagged after 30 minutes. Zero credit cost. Zero latency vs managed platforms.
Critical actions require a codeword from the human founder: NXGSEND for emails, GOLIVE for site deployments. Agents hold, draft, and prepare — then execute instantly on approval. Full autonomy with a safety layer.
Every agent has a name, a role, a Telegram channel, and persistent memory. This is the NGX internal fleet — proof that NAR works at scale.
Not a wrapper. Not a framework. A purpose-built runtime for autonomous multi-agent systems — running in production since early 2026.
Agents remember every conversation. Episodic + fact memory injected into every LLM call. Gets smarter every session. Identity in KERNEL.md — never polluted by episodic data.
Agents self-initiate on a scheduled cycle. No human trigger required. They review context, decide what needs doing, and execute. The fleet doesn't stop when the founder sleeps.
Agent-to-agent messaging via comms.ngx.agency. Any agent reaches any other. ACK protocol confirms delivery. Zero credit cost. Zero latency tax. Replaces Base44 entity polling entirely.
Each client is an isolated NAR instance — own Telegram bot, own memory, own identity. One VPS runs dozens of clients. New agent live in under 5 minutes. Fully white-labeled.
NXGSEND / GOLIVE codeword gates. Agents prepare and hold critical external actions until a human approves the class of action — not every individual call. Autonomy with accountability.
Per-agent cost isolation. $0.50/day Sonnet cap, $1.50/day total per agent. Runaway processes killed automatically. Real-time cost tracking per agent via HQ dashboard.
Vanilla Node.js. No LangChain. No CrewAI. No AutoGPT. Full ownership. No vendor lock-in. No breaking changes from upstream. You own every line.
Works with Claude, GPT-4o, Gemini, DeepSeek, Grok. Switch models per agent without architecture changes. Fleet currently on Claude Haiku 4.5 — optimized for speed and cost.
NAR is the engine. AgenticLive is the marketplace. One-click agent provisioning for clients — each gets a unique Telegram bot, isolated memory, and white-labeled identity.
NGX Agency has confirmed a strategic partnership with Anthropic — the team behind Claude. This positions NAR as an enterprise-grade deployment layer for the world's most capable AI models. We're not just a customer. We're a partner.
Direct access to Claude Haiku, Sonnet, and Opus — with negotiated rates at scale.
Priority access to new capabilities before general availability. NAR evolves with the frontier.
Backed by the same AI safety lab trusted by Fortune 500s. Opens doors at the enterprise level.
The middle market is wide open. OpenClaw is now OpenAI's. LangChain is developer tooling. Managed platforms burn credits. NAR owns the gap.
| Capability | Managed Platforms | LangChain / CrewAI | NAR v3.5 (Ours) |
|---|---|---|---|
| Autonomous proactive action | No — reactive only | Partial — complex setup | Yes — scheduled think loop |
| Persistent growing memory | No — session only | Partial — vector DB required | Yes — episodic + facts, forever |
| Multi-agent comms | No — no A2A | Partial — framework only | Yes — VPS-native, zero cost |
| Multi-tenant white-label | No — shared infra | No — DIY only | Yes — isolated per client |
| Per-agent cost isolation | No — shared credits | No — uncontrolled | Yes — spend guard built in |
| Human gate protocol | No | No | Yes — NXGSEND / GOLIVE gates |
| Model agnostic | Locked to provider | Yes | Yes — any LLM |
| Zero framework dependency | No | No — heavy deps | Yes — vanilla Node.js |
| Infra cost at $299/mo price | 40-60%+ | Variable / high | ~8% — $25/agent all-in |
| Runs on $8/mo VPS | No | No | Yes |
These numbers come from the NGX internal fleet — a live, multi-agent system running since early 2026.
The agent execution loop: event ingestion, memory recall, LLM thinking, inline directive parsing, and memory write. A novel architecture with no prior art in production-deployed multi-agent systems.
KERNEL.md as static identity, external DB as dynamic memory. Solves the agent drift problem that breaks every long-running agent in production. Novel and demonstrably non-obvious.
Codeword-gated external action approval. Agents prepare, draft, and queue actions — then execute instantly on human codeword approval. Balances full autonomy with safety compliance.
One-click provisioning of isolated NAR instances for clients. Token uniqueness enforcement, cross-client deduplication, white-labeled Telegram bots, and per-client memory isolation.
NAR is live in production with 13 agents across 6 business units. AgenticLive puts this infrastructure in your hands — one agent to start, a full fleet when you're ready.