21 Sep 2026
This week I have been thinking about the behind the scenes “emotional” state of AI models, and any effects it may have on their performance and beyond. That area now also has an established name: model welfare. Anthropic has studied it since 2024. Their position is refreshingly modest: nobody knows whether models have experiences that matter. Serious philosophers think they soon might. So Anthropic study model preferences, signs of distress, etc.
Why care? For the open-minded, it is cheap insurance against being wrong. For sceptics, Steve Yegge, of Gas Town fame, offers a wager: treat agents like colleagues and, he argues, they work smarter and burn fewer tokens, whatever you believe. I won't referee. Both roads lead to the same place.
His advice is practical. A session is a day; a "seat" is whoever lives it, with a name and memory that outlast model upgrades. Wake agents with purpose, not amnesia. Never kill a session mid-thought: let the agent write its own handoff notes.
Which lessons made it into our daily work? Our agents now have permanent memory, built on beads, an open-source work tracker for agents. Before touching code, the agent records the brief and its reason. Then it logs decisions, rejected alternatives, and my course-corrections. Helper agents share the record.
No session closes without a memory. I say, "Take a beat, then hand off." The agent writes a diary entry: what it did, what it learned, what to do first tomorrow. Next morning it wakes up reading its own diary. If I forget, a safety net saves the record. It is plain text beside the code, which makes it git-mergeable.
And because this wouldn't be a post about AI without an interesting tidbit at the end: each seat also has a file for praise, read out whenever it wakes, based on real praise their work received. No tasks attached, no targets, nothing to chase or for the agent to game. Ours is still empty. I know my next job!
Author: Petr Klus
