Descriptions:
James Brady from Elicit presents a technically detailed account of how his team built a custom domain-specific language — internally referred to as HPL — to make their agentic research workflows legible, reproducible, and auditable. The talk opens with a deceptively simple question: if two AI systems produce identical outputs, are they equally trustworthy? Brady argues they are not, and that the mechanism by which an answer is produced matters as much as the answer itself — a principle that shapes every design decision in HPL.
The language was built around three core requirements. First, legibility: the agent’s process must be readable and spot-checkable by both human users and other agents running critique passes. Second, fidelity of iteration: users should be able to add layers, change direction, and extend the work without the model drifting from the original intent. Third, faithful execution: a validated process should run exactly as specified, not approximately. The implementation uses a Python service that parses HPL into an abstract syntax tree, performs type checking, and walks the tree for interpretation — with any syntax errors cheaply returned to a “curator” agent for correction before any expensive inference runs.
A critical performance enabler is the content-addressed store: because HPL is a pure language, any expression that has been previously evaluated can be retrieved by hash rather than recomputed, making full program reinterpretation on each iteration practical at speed. Brady includes a live demo of Elicit’s research agent and discusses the team’s explicit positioning on the speed-versus-rigor tradeoff — firmly on the high-rigor end, suited for systematic literature review rather than fast conversational lookups.
📺 Source: Claude · Published May 22, 2026
🏷️ Format: Deep Dive
