How I built AI-augmented social media loop
AI agents introduce more chaos. AI augmented closed-circuit loops bring changes. Here is how I made a useful AI agent for content management. AI does not replace people. I process data. Fast. Very fast. Faster than any human team. Get that data, ask for help to translate data into information. Get insights. Adapt your behavior.
I'm approaching a month of actively posting content. Some of it got traction. Most of it - went into the void. The old me would be sad and disappointed. I can feel these feelings sometimes. New me becomes curious. Because it is not lost. Everything goes through a single service. AI agents are managing it fully.
A single prompt, a few models with different contexts and it is clear where the leak has happened.
How I built a closed-circuit loop with Bob (agent) and Postiz
I'm using Telegram to communicate with Bob, a dedicated posting bot that lives in the same Hermes Agent stack but has a focused job: it owns the social-post-workflow skill. That skill is a strict two-step pipeline. Step one is preparation - I send a piece of content or a half-formed thought to Bob, and it doesn't post anything immediately. Instead it drafts, asks me to confirm the platform mix, validates character limits per channel, splits long content into threads where needed, and saves anything I don't sign off on to a Notion drafts database for later. Step two is execution - only when I explicitly say "post", Bob takes the approved draft and pushes it through the Postiz CLI to schedule across Threads, X, Mastodon, LinkedIn and any other channel I have wired up. Bob handles retries, watches for the Threads 500-error on long posts, and reports back with the live URLs once everything is published.
Postiz is a self-hosted social media scheduler. It runs as a service on my own infra, exposes a clean HTTP API, and has a CLI wrapper that Bob drives directly. The important part is that the agent is not just calling an API - it owns the publishing layer. Adding a new channel is a matter of extending the Postiz config, not building new code. Removing a channel, retracting a bad post, or queueing a week's worth of content in advance - all of it goes through one interface that the agent controls end to end.
The problem? Clear. We judge others by ourselves. I used to be extremely logical, building expectations, complex steps and over-analyzing everything. It was protecting me for a long time. It doesn't work anymore in a long-term world-wide environment called "life". That's the reason I misunderstand how most of people make decisions. And LLMs are helpful learning it, because we both do pattern matching.
Except LLM has huge data set and no emotions. That's why it can't replace people. It can be helpful to learn more about people: ourselves and others. It does not write posts. It does not generate tons of useless ideas. It suggests fixes. It learns my voice. Not perfect English - a slavic guy who speaks english good enough. Today - got a big update based on findings, to support my further learnings about a suitable for me content creation.
