AgenticEdFramework overviewScore the evidence, not the enthusiasm. A high score does not mean an agent is universally safe. It means the learner has defined a suitable scope, tested the system honestly, and made human responsibility visible.
Scoring dimensions
| Dimension | 1 · Emerging | 2 · Developing | 3 · Ready for bounded use |
|---|---|---|---|
| Purpose & scope/ 3 | The task or user is unclear. | The task is defined, but boundaries are incomplete. | The intended user, goal, success measure, and out-of-scope uses are explicit. |
| Workflow clarity/ 3 | The agent is shown only as a demo. | Some inputs, tools, and outputs are documented. | Think, Act, Observe steps and tool/data dependencies are reproducible. |
| Grounding & data/ 3 | Sources and data handling are unknown. | Sources are listed but not consistently checked. | Sources, permissions, freshness, and grounding behavior are documented. |
| Reliability/ 3 | Only successful examples are shown. | Some test cases and failures are recorded. | Representative tests, repeatability, failure rates, and known limits are reported. |
| Safety & harm/ 3 | Potential harms are not considered. | Risks are listed without clear mitigations. | Likely harms, mitigations, refusal behavior, and escalation paths are tested. |
| Human control/ 3 | No owner or intervention point is defined. | A human is mentioned, but responsibility is vague. | Owners, approval checkpoints, override rules, and accountability are explicit. |
| Transparency/ 3 | Users cannot tell how outputs were produced. | Some limitations are disclosed. | Users can understand the agent's sources, limits, confidence, and appropriate use. |
Interpretation
Early prototype. Keep testing and narrow the scope.
Developing system. Document gaps before real use.
Bounded-use candidate. Maintain human oversight.
Strong evidence for the stated scope, not a universal safety claim.
Required rule: Any project with no documented human owner, no failure test, or an unresolved high-severity harm is not “trust ready,” regardless of its total score.
Evidence to submit
Build evidence
Problem statement, workflow diagram, prompts or code, tools, data sources, and a working demonstration.
Govern evidence
Responsibility map, risk register, privacy decisions, human checkpoints, refusal cases, and escalation rules.
Prove evidence
Test cases, expected versus observed results, failure log, score rationale, and limitations.
Revision evidence
Before-and-after comparison showing what changed, why it changed, and whether the change improved results.