Lu Zhang
← Writing
working note

Agentic Scientific Workflows: From Search to Traceable Decisions

A working model for scientific agents that retrieve evidence, plan the next step, preserve uncertainty, and leave a decision trail.

AI agentsscientific workflowRAGevaluation

Scientific work rarely fails because one answer is missing. It usually fails because the path from question to evidence to decision is unclear.

That is why I think scientific agents should be designed as workflow systems, not just chat interfaces. A useful agent should help a scientist or engineer do four things:

  1. Frame the question tightly enough to search.
  2. Retrieve evidence from a known public or internal corpus.
  3. Preserve uncertainty instead of flattening it into confidence.
  4. Leave a trace that another person can inspect.

For AI for Science, the trace matters as much as the answer. If an agent suggests a molecule, a formulation, a paper, or an experimental next step, the user should be able to see where the suggestion came from and what assumptions were used.

My current ScientificLoop direction is built around that pattern: retrieval, tool use, evaluation, and human review. The agent should be able to answer, but it should also route the visitor or researcher to the next useful action: read a source, compare candidates, inspect a trace, or refine the question.

The practical design goal is modest: make scientific navigation less brittle. The long-term goal is stronger: turn repeated expert workflows into evaluated, auditable agent systems.