How SavingsAI Uses OpenClaw + NemoClaw to Build a Multi-Agent Social Media Pipeline

The Problem: Daily Bilingual Content at Scale

Publishing relevant, accurate social media content about UK savings rates every day — in both English and Traditional Chinese — is not a one-person job if done manually. The news cycle moves fast, each article needs two tailored posts for different audiences, and every post needs to be verified live after publishing.

What we needed was an architecture that could handle the mechanical work autonomously while keeping a human in control of the decisions that actually matter. OpenClaw's Telegram-native agent model was the right fit — agents that live in chat threads, respond to messages, and act on your behalf through familiar interfaces, with NemoClaw providing the security and local model layer on top.

Why OpenClaw + NemoClaw on Nvidia DGX Spark?

OpenClaw — taglined "The AI that actually does things" — is an open-source personal AI agent platform that runs locally on your machine and connects to chat apps like Telegram, WhatsApp, and Slack. It provides browser automation, file system access, persistent memory, and a skill/plugin system, and works with Claude, OpenAI, or local models. It is specifically designed so agents can act on your behalf through familiar messaging interfaces rather than requiring a dedicated UI.

NemoClaw is NVIDIA's open-source stack built on top of OpenClaw, adding enterprise-grade security and privacy controls via NVIDIA OpenShell. It runs high-performance open models like NVIDIA Nemotron locally for full privacy, includes a privacy router for connecting to cloud frontier models when needed, and is optimised for NVIDIA hardware — including GeForce RTX, RTX PRO workstations, DGX Station, and DGX Spark. Its tagline: "Deploy more secure, always-on AI assistants with a single command."

We run the full stack on a local Nvidia DGX Spark — Nvidia's compact AI supercomputer built around the GB10 Grace Blackwell Superchip. NemoClaw's guardrails ensure the pipeline behaves predictably even with always-on agents, while DGX Spark provides the inference performance for five concurrent sessions. All model calls, all content drafts, all browser actions happen on-device.

The Architecture: Five Agents, One Telegram Group

Each agent is a separate OpenClaw process with its own Telegram bot account, dedicated topic thread in a private group, and persistent memory file. They communicate via a shared JSON state tracker rather than calling each other directly — this keeps each agent simple, independently restartable, and easy to debug.

How the Pipeline Runs

OpenClaw's built-in cron system fires the Orchestrator each morning. The Orchestrator resets the shared tracker and spawns the News Fetcher. From there, the pipeline moves forward automatically — pausing only at the three human checkpoints before resuming.

Human-in-the-Loop: Three Deliberate Gates

A fully autonomous pipeline that publishes financial content without review would be a mistake. The three approval checkpoints are deliberate — not limitations of the system, but a design choice.

• Article selection: The operator sees what the AI ranked and decides what is actually worth sharing that day. Context, timing, and audience relevance require human judgement.
• Content approval: All bilingual drafts are reviewed before anything is posted. The operator can approve outright or request specific edits — the Content Creator agent rewrites and resubmits.
• Platform choice: Some stories suit one platform better than others. The operator confirms the target for each run.

Everything outside these three gates — news fetching, drafting, posting, QA, rework — runs without any human input.

The QA Loop: OpenClaw Browser Tools in Action

One of the most valuable parts of this OpenClaw setup is using browser automation not just for posting — but for independent verification. The Post Verifier agent has no knowledge of what the Social Poster "thinks" happened. It goes directly to each live post URL and checks three failure modes:

• Content mismatch — wrong article posted, garbled text, or Simplified Chinese instead of Traditional Chinese
• Missing link reply — post is live and correct, but the news article URL reply thread was never added
• Post missing — the URL returns a 404 or blank page

When an issue is found, the verifier passes targeted fix instructions back to the Social Poster and triggers a rework. After the fix, it re-verifies. Only if the problem persists after one rework attempt does it escalate to the operator. In practice, the closed loop handles most issues without any manual step.

Keeping Agents Alive: The Heartbeat Pattern

A practical challenge with long-running OpenClaw sessions: agents can naturally treat the end of each task unit as a stopping point, especially when posting five or more articles sequentially. The Social Poster needs to stay active across multiple articles and platforms without the operator sending manual "continue" messages.

The solution is a lightweight heartbeat — a short-interval OpenClaw cron job that reads the shared state tracker during active hours. If the posting step is still running, it sends a quiet nudge to the poster session to continue with the next article. If everything is done, it does nothing. This pattern works cleanly with OpenClaw's session model and removes the need for any manual babysitting of long runs.

Persistent Agent Memory: Learning from Feedback

Each OpenClaw agent has its own persistent memory file — a writable text store that survives across sessions. When tone guidance changes, a new posting rule is added, or an agent is corrected on behaviour, it writes the instruction to memory immediately. On every subsequent session start, it reads the file back.

This means the agents improve continuously through natural Telegram conversation feedback — no redeployment, no prompt engineering sessions. A single correction becomes a permanent behaviour change. Over time, the agents develop a detailed operational knowledge base that reflects the specific requirements of the pipeline.

Tech Stack Summary