Coding an Agent: Steering a Local Model

Last Updated on October 6, 2026 by Editorial Team Author(s): Enzo Lombardi Originally published on Towards AI. Directional Edits in DS4, or why a slider can do what a fork of the weights cannot Most of what you change about a language model’s behaviour you change from the outside. You write a system prompt, you pick a sampling temperature, you ask for brevity in the last line of the message. The model reads those words like everything else and decides how much to care. There is a second lever that most people never touch because most runtimes do not expose it: you can reach inside the forward pass and edit the activations directly, one layer at a time, while the model is running. The article explains DS4 activation steering—using per-layer direction vectors and a scale during inference—to shape model behavior without fine-tuning or loading new weights, including the mechanics of how projections are applied in attention/FFN outputs and what trade-offs (like decode speed) come with it. It then walks through how a steering direction is built from contrasting prompt lists (e.g., succinct vs verbose), demonstrates the effect with an agent-harnessed example measuring verbosity changes, …

≈