Duolingo is centralizing how it builds and scales autonomous AI agents to make language lessons more personalized. This shift toward automated tutoring means parents and educators must balance technological changes with student privacy and teaching quality.
What Happened
According to a technical update from the Duolingo engineering team, the company built a shared platform to standardize how its developers deploy autonomous AI agents. Previously, engineering teams had to rebuild infrastructure and write unique code for every project. Now, Duolingo separates the agent's definition, like prompts and tools, from its execution using a workflow engine called Temporal.
Developers can register an agent once and run it across internal sites, Slack, and the Duolingo app. The platform supports multiple models, including the Claude Agents SDK and OpenAI Agents SDK. As we previously reported, companies are investing heavily in agentic education systems to move beyond pre-programmed lessons.
The Bigger Picture
While autonomous AI agents promise customized tutoring, they also introduce teaching risks. Unlike static software, an agent responds to student input dynamically. This freedom often leads to instructional failures.
According to a benchmark study called SafeTutors, AI tutoring systems frequently suffer from a "quiet erosion of learning" where they give away answers instead of guiding students step-by-step. The study found that while AI agents can appear safe in short exchanges, their teaching failure rate spikes from 17.7% to 77.8% during longer conversations.
Students are also skilled at bypassing educational guardrails. A paper on the SHAPE benchmark shows that students manipulate models using specific prompts to force direct answers. Without strict boundaries, autonomous agents can stall academic progress. These systems must rely on frameworks like self-adapting skills and specialized memory architectures to remember past mistakes and adapt rather than giving away the solutions.
What This Means for Families
Duolingo’s architectural shift highlights legal and privacy issues. When an educational platform routes student input to external model providers, it triggers strict regulations. Under the Children's Online Privacy Protection Act (COPPA), platforms must list all third-party AI vendors as subprocessors in their privacy policies and notify parents directly before collecting data on children under 13.
In classrooms, the Family Educational Rights and Privacy Act (FERPA) dictates that sharing student personally identifiable information with third-party AI tools is only legal under the "school-official exception". This exception requires the vendor to remain under the direct control of the school district. Standard consumer AI terms also allow providers to train models on user inputs. To protect student privacy, schools and platforms must establish strict enterprise-level Data Processing Agreements that forbid downstream model training.
What You Can Do
Parents can regularly check their child's learning history to ensure the AI tutor prompts them to think instead of feeding them correct answers.
Look for explicit disclosures of third-party AI subprocessors in privacy policies to understand where student data goes and whether it is used for model training.
Educators can evaluate tools using the UK Department for Education product safety standards, which offer a 13-question framework to audit AI safety and content loops in schools.