Do LLM Agents Still Need to Be Told How to Self-Improve?
A gameplay study finds capable agents can devise their own learning loops from basic tools and a goal, then transfer some learned text artifacts to other tasks. It tests an early form of agent-directed improvement.
A frozen agent received a compute budget, experience-gathering tools, editable text documents, and a general self-improvement goal. It experimented without a prescribed learning loop or real-world data, and some learned artifacts transferred across game horizons, mechanics, and 3D navigation.
For coding agents, this suggests testing whether they can improve task-specific notes or procedures when given clear goals, tools, and bounded resources. Measure transfer to held-out tasks rather than assuming a self-written playbook will generalize.
A frozen agent received a compute budget, experience-gathering tools, editable text documents, and a general self-improvement goal. It experimented without a prescribed learning loop or real-world data, and some learned artifacts transferred across game horizons, mechanics, and 3D navigation. For coding agents, this suggests testing whether they can improve task-specific notes or procedures when given clear goals, tools, and bounded resources. Measure transfer to held-out tasks rather than assuming a self-written playbook will generalize. The evidence comes from gameplay, and the material gives no quantitative effect sizes or cost comparisons. It does not establish that the approach works for software engineering tasks.
This makes self-improvement an agent capability to test: a frozen agent can use tools and editable artifacts to discover procedures without a prescribed learning loop, with some transfer across gameplay settings. It invites coding-agent experiments with bounded resources and held-out tasks, while leaving software-engineering transfer and quantitative benefit unestablished.