---
title: "Agents that almost work"
date: "2023-04-27"
summary: "Everyone's passing AutoGPT around like a magic trick — I gave one a real job and watched it unravel, and the failure points straight at what's missing"
tags: ["ai","agents","autonomy","grounding"]
---

A friend sent me a screen recording last week: an agent he'd pointed at a vague goal, churning through step after step on its own. "Watch this," he wrote. I watched. For about four minutes it was the most exciting thing I'd seen since December.

Then it started talking to itself. A plan, a sub-plan, a sub-sub-plan. A task that spawned three tasks that spawned a question it had already answered two steps earlier. It looped. It congratulated itself for finishing work it never did.

AutoGPT, BabyAGI — everyone is passing them around like a magic trick. Give it an objective, walk away, come back to a finished product. The dream is older than the tools. I have wanted exactly this for years.

So I spun one up myself and gave it a real job from the studio. It opened with confidence and unraveled by the tenth move. Not because it was stupid. Because it had nothing to hold onto. No memory of what it tried. No ground truth to check itself against. It was reasoning in a vacuum, and a vacuum is where good reasoning goes to spin.

That is the part nobody is posting about. The failure is the lesson. Intelligence on its own drifts. What an agent lacks isn't cleverness — it's footing. Something fixed and trusted to stand on while it thinks. Facts it can return to. Steps it can verify instead of imagine.

We are all staring at the autonomy and missing the floor underneath it. Make the goal too big and the model dreams; give it solid ground and it walks.

The loops will get fixed. I am more interested in what they're loudly telling us we don't have yet.