Reading Group: AutoHarness: Improving LLM Agents by Automatically Synthesizing a Code Harness

​Join the Snorkel AI Reading Group, a recurring forum to explore the latest frontier developments in AI while building meaningful connections within the community.

In this session, Carter Wendelken of Google DeepMind will dive into his paper “AutoHarness: Improving LLM Agents by Automatically Synthesizing a Code Harness."

Agenda:

​5:30pm - doors open
6pm - talk begins
Light drinks and appetizers provided

​Most work on language models focuses on improving the model itself. This talk takes a different angle: improving outcomes by shaping how the model acts through automatically generated code harnesses.

​You’ll learn:

• ​Why LLM agents often fail when their actions are prohibited by the external environment

• ​How AutoHarness automatically synthesizes a code harness through iterative refinement from environment feedback

• ​How a learned harness can prevent illegal moves in structured environments

• ​How execution-guided search over code can yield a more robust control loop

• ​Why a smaller model with a custom harness can outperform larger models

• ​How the agent can generate the entire policy in code, removing the need for model inference at decision time