Hyperagents

A hyperagent bundles the agent that solves the task and the meta agent that edits both itself and the task agent into a single program, so the procedure that generates improvements is itself open to modification. The reference implementation, DGM-Hyperagents, extends the Darwin Godel Machine’s generate-and-evaluate loop beyond coding to any computable task, and its meta-level gains transfer across domains and accumulate across runs (arXiv; code on GitHub). In this corpus it is the system behind hyperagents edit their own mechanism, the formal counterpart to ALMA’s meta-learned memory.