AI Tutor Team Details Real-Time Teaching Architecture for Ages 4-9, Drops Standard Tool Loop

An AI tutor team says it set out to build the first AI tutor teaching math and reading to children ages 4-9, and shared the architectural decisions behind it. The team notes frontier models take 2-3 seconds to first token and decode at about 30 tokens per second, so a standard tool loop creates 3-4 seconds of downtime between each sentence. It dropped the standard agent loop, built a custom harness that streams multiple actions in one response, and split the system into a converser and an asynchronous planner.

This is a clear trade-off by an AI education product in a real-time teaching setting against general agent frameworks: the team says frontier models are heavily post-trained on the tool-use pattern, while real-time learning sits at the other extreme of the speed-versus-thinking trade-off, so it owns the loop and builds its own observability and tracing. For AI tutor products aimed at young children, latency directly affects whether children stay engaged, and the playtest observations and architecture choices offer a verifiable engineering reference for similar products. The limitation is that the post provides no quantitative learning outcomes, user scale, or independent evaluation; all conclusions come from the team itself. AI tutor products for young children need to treat response latency as a core design constraint, and the standard tool loop may not fit Building a custom harness and asynchronous planner adds engineering and observability costs, and the team says it may switch back to a simpler loop if future models get fast enough