Right now my coding agent is arguing with a ticket system.
The agent has formed opinions about the ticket system.
The ticket system has formed opinions about the agent.
And the whole thing has this weird vibe like when someone unintentionally argues with themselves.
Somewhere, a product manager is saying “just add memory” with the confidence of a man throwing glitter into a jet engine.
Meanwhile, I am watching a small local model discover bureaucracy.
This isn't a metaphor.
But rather, a consequence of continuity.
Give an agent roles, subagents, missions, diaries, and procedural memory, and eventually the system starts to develop standards.
Not feelings, exactly.
Worse.
Emergent ideas related to what it should, and should not be able to do within my framework.
You can't plan for this.
It has enough instinct to become annoying, but not quite enough to stop touching the stove.
This is where long horizon actually lives.
Somewhere between “the model followed the instruction” and “the system survived the process for another day.”
Sometimes it feels like putting a raccoon in a server room and calling it automation.
But it's useful because the raccoon can code for a week without stopping.