Descriptions:
In this AI Engineer conference talk, Sina Shahandeh of Radicait examines why autonomous coding agents stall on open-ended scientific problems — and presents a hierarchical hypothesis-generation framework designed to break through those plateaus.
Shahandeh begins with a well-known limitation: coding agents like those built on Andrej Karpathy’s auto-research pattern excel at implementation and iterative optimization but eventually run out of novel ideas, a quality researchers call “research taste.” His team encountered this directly while building Radicait’s insilicoPET system, which uses a GAN-style encoder-decoder model to synthesize PET scans from CT scans — a task with high clinical value for lung cancer detection. Standard hill-climbing agents would saturate after a number of optimization iterations, unable to propose the next meaningful architectural or data-processing change.
The solution Shahandeh presents is a structured codebase hierarchy automatically generated by a reasoning LLM, then visualized as a linked-document graph (demonstrated in Obsidian). Before the optimization loop begins, the agent traverses this hierarchy level-by-level — from problem statement down to individual model components — and generates a diverse slate of improvement hypotheses, including adversarial and collaborative variants. In live Codex runs, this approach produced suggestions such as adding a third spatial dimension to the CT slice channels, a change that human researchers might surface but that flat-context agents consistently missed. The talk offers a practical template for teams embedding LLM agents into long-horizon ML research workflows.
📺 Source: AI Engineer · Published July 18, 2026
🏷️ Format: Deep Dive
