A paper on arXiv introduces SpecForge, an approach for coding agents built on frozen language models. The key idea is that these agents can improve their performance by refining the external context they rely on, without updating the model weights. The authors frame this as a study of recursive self-improvement through behavioral specifications.
Behavioral specifications appear to act as a form of external guidance that the agent can adjust over time. Because the underlying model stays frozen, the improvement comes from the agent's ability to modify what it uses or sees, rather than from retraining.
The abstract is brief, but it points to a research direction where self-improvement is achieved through context engineering. This could make coding agents more adaptable without the cost of fine-tuning, though the paper's details and results are not yet available in the provided summary.