For serious developers or engineers
by•
I have been using Claude code and cursor for the past almost 2 years. Everybody is talking about loop engineering and graphs. I have such a setup with over 10 agents. It used to work pretty well.
Over the past month, the more autonomy I give to the agents the more complicated they make the solution thus getting trapped in a loop for hours. They tend to find bugs and add nice to have features that are not even nice to have when the effort is huge and the value is low.
Am I the only one experiencing this strange behavior?
28 views
Replies
Would have to know your setup (tools, models, effort). I’ve experienced something similar through a large multi-agent system. More autonomy isn’t equivalent to better results, at some point they begin increasing the problem vs. solving it.
For example, in my system I’ve moved toward deterministic agent contracts and fail-fast behavior rather than allowing agents to always improvise. Each agent has a set responsibility and execution contract. A slew of agents, or how I classify them “Super Agents”, are good, but they will find things that you weren’t looking for without implementing the proper boundaries.
I would push back a bit on the "built to burn tokens" theory though. What I see is simpler: nothing in the loop tells the agent it's finished, so it keeps finding work
What fixed most of it for me was making "done" something a test can check. I keep a rules file with a short Never list, and the rules I really care about are pinned by tests. One fails the build if a marketing claim we retired shows up again anywhere in the app copy. The agent doesn't have to remember the rule. It can't get green without keeping it.
The other one is a standing instruction that the smallest diff wins, and anything nobody asked for gets named in one line instead of built. That cut most of the "nice to have" features
Their architecture is designed to keep you longer and burn more tokens.
I don't like this mode; it just wastes tokens, and sometimes the solution it comes up with is wrong.