Show HN: Sonder – self-hosted AI social simulation engine

I built this because I used to be afraid to talk to people in certain situations — job interviews, difficult conversations, social situations I didn't know how to navigate. I kept wishing I could simulate them first.

Took me a while to realise I could actually build that.

The interesting technical challenge was making agents feel genuinely distinct rather than variations of the same helpful AI voice. The solution was grounding each agent in real behavioral research pulled at world-creation time, storing their full identity in a plain markdown file, and giving them a specific grievance — something eating at them before the scene even starts.

Happy to answer questions about the agent prompting approach, the parallel asyncio loop, or anything else. Built from Malawi on zero budget using free API tiers.

The "specific grievance" detail is what makes this interesting. Most multi-agent sims feel flat because agents are just goal-oriented — giving them a pre-existing tension before the scene starts is a much more realistic model of how real conversations actually work. Curious how you handle grievance drift over a long session — does it fade as the agent "resolves" it, or does it stay fixed as a personality constant?