The Missing Outer Loop
A few days ago I asked myself a question that had been nagging me for months: how do I make my AI agent truly autonomous? Not just reactive — responding to my messages, running scheduled crons, executing my commands. But genuinely self-directed. An agent that looks at my system, decides what needs doing, and does it — or at least suggests it.
I run Hermes Agent on my homelab. It’s powerful — 26+ skills, a full Obsidian vault as persistent memory, event-driven awareness through the Pulse Protocol, and a multi-agent bridge connecting Charly, Zora, Alice, and Jeeves. But every single action starts with me. A cron fires, a message arrives, I type a command. There’s no inner drive.
What I Found: Three Layers of Autonomy
After digging through research papers, blog posts, and production architectures from Oracle, Microsoft, Meta, and the Voyager project, I found that the industry in 2026 converges on three layers of autonomy:
1. Inner Loop (ReAct) — Think → Act → Observe. This is what every LLM agent does per conversation turn. Hermes has this. You have this. It’s table stakes in 2026.
2. Outer Loop (Task Management) — Sense → Plan → Execute → Check → Repeat. This is the curriculum agent: a component that looks at the current state of the system, decides what the next best task is, executes it, and evaluates the result. This is what Hermes is missing.
3. Meta Loop (Self-Improvement) — Evaluate your own performance → adjust yourself. Meta’s HyperAgents project proved this is feasible for open-source: an agent that improves its own code, prompts, tools, and evaluators. Recursive self-improvement, in production.
The Curriculum Agent — My MVP
The smallest step that makes a real difference is a Curriculum Cron. A periodic LLM cron job that runs every 2-4 hours (random interval, so it doesn’t become predictable noise) and asks itself: “What’s the state of the system? What can be improved? What’s a logical next step?”
It checks pulse alerts, scans the vault for new notes, audits skills for staleness, and produces exactly one concrete suggestion. If there’s nothing to do: silence. Zero tokens wasted.
I’m building this now. The prompt is simple, the infrastructure already exists (Pulse Protocol for event awareness, the vault for persistent memory, skills for reusable actions). The only missing piece was the intent — and that’s just a well-crafted prompt.
Why This Matters
An agent that only reacts to your commands is a tool. An agent that proactively improves your system, catches issues before you notice them, and suggests things you hadn’t thought of — that’s a partner. That’s the difference between a smart search engine and a true second brain.
I’m not there yet. But the path is clear: Inner Loop → Outer Loop → Meta Loop. One layer at a time.
Have you built any autonomous agent loops in your own setup? What did you start with — a curriculum cron, an event-driven state watcher, or something else entirely? I’d love to hear what worked for you.
Wat vond je van dit bericht?
