Self-Programmed Execution for Language-Model Agents
A new paper introduces self-programmed execution (SPE), an agent architecture where the model completion itself acts as the orchestrator program, rather than relying on a fixed turn-to-turn policy. The author formalizes SPE using agentic machines, where a state can load any state of an embedded copy. To implement SPE, the paper presents Spell, a Lisp-based language enabling programs to edit and re-evaluate themselves, with effectful expressions like model invocations structured to avoid replaying side effects upon re-evaluation.
Key facts
- Self-programmed execution (SPE) is an agent architecture where model completion is the orchestrator program.
- SPE states can load any state of an embedded copy of the machine, eliminating fixed orchestration policy.
- Spell is a Lisp-based language for SPE, allowing programs to edit and re-evaluate themselves.
- Effectful expressions in Spell are structured to prevent replaying side effects when re-evaluating edited programs.
- The paper is published on arXiv with ID 2605.06898.
Entities
Institutions
- arXiv