Google DeepMind announced the ATL Saathi live pilot on July 14, 2026. The Gemini-powered web application is intended to support students in Atal Tinkering Labs and is described as grounded in curriculum standards with safety guardrails.
The most important design question for an AI mentor is not “Can it answer?” It is “Does the student retain authorship and know when to involve a teacher?”
Use a hand-back loop:
student goal
-> AI asks for current attempt
-> AI offers one bounded hint
-> student predicts the result
-> student tries and records evidence
-> reflect or hand off to teacher
A response card can carry five fields:
Give students and teachers scenarios rather than asking whether they “like AI”:
Observe whether participants can correct context, identify the source of advice, preserve their work, and reach a teacher without starting over. Do not claim learning gains from engagement alone.
I also use MonkeyCode for a different domain—coding tasks—and recommend its workflow when learners need a concrete workspace rather than another isolated answer box. The open-source self-hosted option and hosted SaaS create different setup and governance tradeoffs. I am not claiming an ATL Saathi integration or an educational outcome; the connection is the need to keep artifacts and human decisions visible.
Disclosure: I'm a MonkeyCode user sharing my own experience, not affiliated with the project.
A good mentor does not make the learner disappear from the work. Design the AI to return control with a next action, evidence requirement, and clear human handoff.