A wrong answer and a misunderstanding are different findings
Two students receive the same mark on the same question for unrelated reasons. One has misread the instructions. The other does not understand the concept the question assesses. Feedback that treats both cases as identical helps neither student.
You tell them apart by looking at the rest of the work. A concept gap leaves a trail: the same assumption repeated three times, and a definition used in the wrong sense throughout. Carelessness occurs once and is then contradicted somewhere else in the same submission.
This distinction affects what the educator does next, so the agent is asked to make it rather than assign a score. The agent provides evidence and an interpretation. Assessment stays with the course teacher.
The agent does not give grades, and the material does not go directly to the student. The output is a draft, with the passages it is based on, for a teacher to edit, disagree with, or discard.
How the agent is built in Agent Studio
Your prompt should include the rubric, your department's feedback register, and, most usefully, previous submissions with the errors you have already marked as conceptual or as slips.
The system prompt makes it clear that the agent gives no marks, assigns no grades, makes no comparisons with other students, and includes no statements about ability. Findings refer to a single piece of work only.
Each claim must come with the passage it is based on, so an educator can check the reading at once instead of rereading the entire assignment.
For each criterion in a rubric, you need three states: met with evidence, not met with evidence, and insufficient evidence in the submission. The third state is the one where the marker has to make a genuine judgment.
The marking flow passes the extracted submission to the Agent Chat node. Access is limited to the assignment, not to the student's record or their history from other courses.
Google Docs receives the draft, and the educator edits it there. A Human in the Loop node holds the release. Nothing reaches the student until this node has been passed.
Airtable records instances where the educator disagreed with the reading. Those disagreements develop prompt skills, so the agent learns your particular standards rather than general ones.
What this agent is built from
- Prompt skills: rubric, department feedback register, and past submissions with errors already classified.
- System prompt: no mark, no grade, no comparison between students, and no claim about ability.
- System prompt: every finding quotes the passage it rests on.
- Agent Chat (Util): receives the extracted submission, scoped to this assignment alone.
- Google Docs (Integration): holds the draft that the educator edits.
- Email (Human in the Loop): gates release. Nothing reaches a student ungated.
- When an ActionFlow is enough: If checking a submission against rubric criteria and quoting the matching passages is the work, build the workflow. That sequence covers straightforward marking.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.