Descriptions:
Aditya Bhargava, staff engineer at Etsy and author of the popular book “Grokking Algorithms,” makes a counterintuitive case at the AI Engineer conference: the harness surrounding a language model matters more than the model itself. He cites HarnessBench, a new benchmark covering 106 standardized tasks, which shows that swapping only the harness while holding the model and task constant produces accuracy swings from 52.4% to 76.2% — a gap of more than 20 percentage points. Notably, weaker models benefit the most from a superior harness, which Bhargava frames as liberating: developers can achieve frontier-level performance using locally runnable open-source models if they invest in harness quality rather than paying for proprietary APIs.
To pursue this thesis, Bhargava spent six months building Agency, a new programming language designed specifically for constructing agent harnesses. Agency provides built-in primitives that every agent needs: automatic interrupt signals on destructive or sensitive tool calls, partial function application to lock LLM-accessible parameters to safe scopes, and structured human-approval flows that balance safety with autonomy.
The second half of the talk is a live walkthrough of a coding agent built in Agency, evolved across seven progressive iterations. Each version uses the same model and the same task, but an improved harness raises performance at every step. The demo covers raw model invocation, tool attachment, safe interrupt handlers, and scoped directory access — making it a concrete, reproducible guide for any engineer looking to extract more capability from smaller, self-hosted models.
📺 Source: AI Engineer · Published July 07, 2026
🏷️ Format: Hands On Build
