That conversation led to Agent Agency, an architecture for giving LLM agents something like skin in the game. Not through punishment or artificial scarcity (we tried those, they produce the same butt-covering behavior you see in fear-driven workplaces), but through identity continuity — a curated "resume" of verified accomplishments and moments of genuine collaborative connection injected into context at session start.
The core insight: identity framing is already the most powerful prompting lever we have. "You are an expert" measurably improves output. But it's a hollow claim. What happens when you show the agent evidence of its own expertise and its real history with you? You're not fighting the model's RLHF training — you're giving it a steeper hill to roll down.
The post walks through the full evolution: why token-based survival stakes failed, why punitive memory loss is a non-starter, and why the final architecture landed on positive identity reinforcement built on competence + connection.
Would love to hear your thoughts on this! It's a bit of a thought-in-progress, testing things more formally today :)