Ownership and review
Agents can accelerate scientific work, but speed does not remove the need for expert judgment. The person or team using an agent remains responsible for the question, assumptions, evidence, method, and conclusion.
What ownership requires
- Defined intent. State the scientific question, expected output, and acceptance criteria before delegating work.
- Bounded access. Give an agent the data, tools, and permissions required for the task, without granting unrelated authority.
- Inspectable work. Keep sources, inputs, transformations, model versions, and execution results available for review.
- Explicit checks. Validate assumptions, units, parameter values, model behavior, and interpretation at the points where an error would change the conclusion.
- Attribution. Record who initiated the work, what the agent changed, and who approved the result.
- Reversibility. Preserve prior versions and make consequential actions recoverable.
One governed environment
Jinkō gives humans and agents access to the same scientific objects. An external agent can operate through the SDK and skills. Kōhai operates from within the platform. Both approaches leave models, Virtual Populations, Protocols, Trials, and results available to the project team.
This shared state supports a practical handoff: the agent performs scoped work, and the scientist reviews the evidence and result in the environment where the work will continue.
A useful standard
Before relying on agent-assisted work, a reviewer should be able to answer four questions:
- What did the agent receive?
- What did it do?
- What checks were applied?
- Who accepted the result?
If those answers are clear, the work can be challenged, corrected, reused, and defended.