Making agentic workflows trustworthy and verifiable with a custom DSL
- Source
- youtube.com
- Author
- Claude
- Date
Context
Elicit, a research-assistant product, has its AI agent write its plan in a restricted programming language so that people can inspect the steps. In a talk on Claude's channel, Elicit's James Brady says two systems can give the same answer yet deserve different trust depending on how they got there, so the should be legible, stay faithful across edits, and actually be followed.
Elicit's language, called Ash PL, is a typed, purely functional subset of Python with no loops, recursion or mutation, plus built-in operations for retrieving papers and clinical trials. A writes the , ordinary Python code interprets it, and the model rewrites it from the results, for example fixing a type error and rerunning. Brady shows an example workflow for a competitive analysis with web searches, joins and enrichment steps.
Brady describes this as the team's own design choice and says it is not for everyone. The excerpt gives no measurements of whether it improves trust or accuracy. The model-facing writer is swappable, and the team currently uses the Pi with Anthropic models.
- Process — A running instance of a program, with its own working memory, open files and identity.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Program — A set of instructions stored so that a computer can run it.
- Harness — The software that sends a model its context, recognizes tool calls, runs or refuses them, returns results and keeps the record.
Checking sign-in…
Loading comments…





