What makes an agentic move worth using?
An agentic move is a repeatable educational workflow: an agent pursues a goal, uses tools or makes routing decisions, checks its result, and hands control to a person at defined checkpoints.
We select the workflow in its context. A new tool announcement starts research; it does not earn a recommendation. Version 1.0, effective October 18, 2026. This is AgenticED’s editorial methodology, not a validated rating scale or school adoption approval.
Eight gates. Every one matters.
Each gate receives Pass, Unknown or Fail with evidence and reasoning. Unknown holds a candidate for more evidence. A failed gate prevents selection in that form. There is no average score that can hide a privacy, accessibility or human-control failure.
- A worthwhile educational purpose. Name the educator or learner problem, intended users, context, and a useful outcome. Compare against a simpler manual or non-agentic approach.
- A bounded agentic workflow. Identify the goal, tool use or routing decision, result check and stop condition. A product launch or one-off prompt alone is not an agentic move.
- Preserved learner thinking. Specify the reasoning, practice, retrieval or judgment the learner still performs. Do not substitute an agent’s completed work for the learning objective.
- Source support and a reproducible test. Verify capabilities against cited sources and document a normal case, an error case and a stop/escalation case using public or synthetic inputs. Separate technical success from learning effectiveness.
- Appropriate data and permissions. Use only public or synthetic materials in AgenticED. Record source rights, minimal permissions and local authorization requirements. Protected records and covert monitoring fail this gate.
- Accessible participation. Check the actual workflow and output for the intended users, including keyboard, assistive technology, language and device needs, or provide a tested equivalent route.
- Meaningful human control. Identify who reviews what before external sending, publishing or changes to authoritative records. Define failure handling and a practical fallback. Exclude automated consequential student decisions.
- Feasible operation over time. Record model and tool versions, permissions, cost and maintenance limits, required staff skills, and measurable success criteria. There must be a responsible owner and re-review date.
How evidence changes the decision
| Evidence available | What it can support |
|---|---|
| Documentation only | Candidate research only; no selected move. |
| Recorded public/synthetic replay | Conditional guidance after all gates pass and an editor verifies a recorded replay. |
| Educator pilot | Recommended guidance can be considered within the tested population and conditions; a pilot does not establish causal learning gains. |
| Relevant independent study | Learning-outcome claims require editor-verified, relevant study evidence and clearly stated limits; this label is not itself a causal evidence rating. |
Every candidate needs a normal-task test, an error or conflicting-evidence test, and a permission-boundary or stop test. The reviewer records the setup, model/tool versions, observed results, restrictions and a report link. A successful replay demonstrates behavior under those conditions. It does not demonstrate improved learning.
Examples of the distinction
- Candidate agentic move: inspect public course-resource links, check responses, flag ambiguous failures and prepare a repair plan for an educator to approve.
- Candidate educator workflow: retrieve from an approved lesson source, draft practice questions, check answer support, flag uncertainty and stop for teacher review while preserving the learner’s problem solving.
- Assisted task: a single prompt that writes a lesson plan. Useful assistance may be possible, but that alone does not demonstrate an agentic workflow.
- Inappropriate move: ingest protected learner records into the public service and make placement or discipline decisions automatically.
These examples illustrate the methodology. They have not been awarded selection status by appearing here.
How Grokbot will keep it current
Grokbot can discover public evidence, deduplicate candidates, propose updates and flag changed or unsafe practices. An editor verifies the sources and test report, reviews every gate, and approves the exact draft. Grokbot cannot certify its own review or publish its own proposal.
Selection expires 30 days after the recorded test. New evidence snapshots with changed source hashes, a methodology-version change or an editorial withdrawal remove a move from the current selection catalog. Agents must check that catalog before adopting a practice. Changes in model capability, permissions, terms, cost, accessibility or observed behavior require an earlier review.
The existing discovery and weekly source schedules remain disabled until deployment and integration are verified. Grokbot is not connected by this update. External agents receive advice; they do not change their own configuration automatically.
Currently selected moves
No workflows have completed this methodology yet. Existing starter tools and guidance remain reference material; they were not automatically promoted into the selected catalog.
The same rules for people and agents
Read the methodology as JSON · Read the current selection catalog. The read-only MCP service exposes both resources.
Research informing the approach
NIST’s AI Risk Management Framework informs our attention to context, oversight, testing and ongoing review. UNESCO’s guidance on AI in education informs the emphasis on human agency, inclusion and educational suitability. The eight gates, evidence thresholds and 30-day review rule are AgenticED’s own policy choices.