How Agents Run the Research Pipeline
How research rules become contracts that agents can execute, why every rejection is a first-class output, and which decisions must stay with people.
Key ideas
Agents run pre-registered gates with frozen thresholds; a step belongs to agents only if its correctness can be verified without asking the executor's intent.
The ledger of rejected candidates, kept append-only with its config hash and data vintage, is a first-class output that stops agents from re-mining falsified directions.
Admission to the live book, risk posture, and opening or closing research lines stay with a human, and the research agent never has a path to place orders.
Read the full essay
Written by Wayland Zhang, founder of Dnalyaw.