Headlong: a microharness for persistent agents

Why we picked it · the editor's summary

Headlong is a loop: call the model, ask it to choose the next thought given its past thoughts, append, repeat. Observations from a user or the environment are injected into that single stream, and the agent decides whether and when to reply, which is what the authors mean by persistent agency. Their shared instance, Audel, folds several users' conversations into one trajectory with memory that decays in resolution over time, and in one episode it diagnosed and repaired its own recall process over 48 minutes with nobody directing it. The authors say plainly that they have no quantitative measure of what persistence is worth and ask for help building one. It is alpha research software from Laude and MIT: give it a spend- capped key and a Docker sandbox before you let it think overnight.