As so many others, I played around with OpenClaw (https://openclaw.ai/). I guess I did not find any revolutionary, but as always, I tried to sketch up how 'my' agent works. So, here we go.

This is an autonomous daily briefing and social posting agent called Owlie42. Owlie42 wakes up at 6:00, reads and processes news using a local LLM (GLM-4.7 via Ollama), stores relevant context in a memory file system, and posts a curated summary to Discord by 6:15. Meanwhile, a background heartbeat continuously monitors bulletin boards and keeps the memory store fresh. It's a lean, fully local AI pipeline with no cloud LLM dependency — just scheduled jobs, file-based memory, and a Discord webhook as the output channel.

Architecture Overview: Owlie's Automated Intelligence System

The diagram shows a multi-layered automation architecture centered around an AI agent named Owlie42, who operates through OpenClaw with a local LLM backend via Ollama running GLM-4.7 on a hardware named Ada.

(Yellow) Business Layer

At the top, three scheduled triggers drive the entire system. The first is a recurring every-30-minutes timer that continuously monitors bulletin boards — likely some form of message board or forum scraper. The second and third are time-based cron triggers firing at 6:00 and 6:15 respectively. The 6:00 job initiates a "read news" workflow, while the 6:15 job kicks off a "Post to Discord" workflow. Both feed into a shared Cron Jobs component, which acts as the central scheduler for the morning routines.

(Blue) Application Layer

The blue components represent the active runtime processes. The Heartbeat mechanism — triggered by the 30-minute timer — runs a continuous health-check or polling loop. It writes state or observations into a MEMORY.md file, functioning as a persistent short-term memory store for the agent.

The Daily Reading workflow is responsible for ingesting and processing the morning news. It reads from a /memory/xyz. directory, which appears to serve as a shared memory space accessible by multiple components. Interestingly, Daily Reading also has a dashed feedback arrow pointing back to MEMORY.md, suggesting it enriches the memory file with newly processed information.

Daily Discord takes processed content and pushes it out via the Discord API. There's a dashed line from Daily Discord back to Daily Reading, hinting at some coordination or content handoff between the two workflows before posting.

(Green) Infrastructure Layer

Ollama is the local LLM runtime, pulling model weights from GLM-4.7 (shown as a data/file artifact). Ada calls Ollama directly for language model inference.

OpenClaw is the agent orchestration framework. It connects upward to the blue process layer — specifically feeding results into Daily Reading and receiving instructions from it. OpenClaw also has access to the markdown files in folder /memory/, enabling it to read and write contextual memory.

The Discord API sits slightly apart, it is the only component not running locally.

Of course, in a more production-like environment, the setup would a lot different. We published a more elaborate piece on this topic at https://desrist2026.org/, see https://rachmann-alexander.github.io/businesscard/res/desrist2026.pdf for the poster.