Keep answers fast

Studium grades an answer and chooses the next screen in a few milliseconds. What you wait for is the assistant that relays it. Three choices decide most of that wait:

  • A session for the lesson. Play a lesson in a new, short session, not at the end of a long working conversation: every answer carries the whole conversation behind it.
  • A fast model. Relaying needs no reasoning, since Studium grades and routes. Keep the larger model for a question that needs a real explanation.
  • The MCP connection described in Connect your assistant, not shell calls to the HTTP API: each shell call starts a process.

What was measured

Measured on the demo lesson atomistique-l1 with tools/latency_bench.py (Claude Code, medians over six answers, 6 November 2026):

Who relays End of the reply First word
a large model, through shell calls, at the end of a long working session (a real trial, desktop app) 27 s about 20 s
the same large model, short session, MCP tools 8.2 s 5.4 s
the fast model, short session, MCP tools 5.4 s 2.0 s
the lesson agent studium-lesson as the session (see Claude Code) 5.5 s 1.9 s
a large model handing every answer to that fast agent 16.1 s 13.5 s

One run per line: a difference of a second between two lines is not established. The bench measures a relay in Claude Code; it does not measure the other clients on this page, or what anyone learns. Lowering the reasoning effort changed nothing measurable. The last line is the trap: an agent placed between you and the server makes two models write every screen.

Play in forms

A prototype goes one step further: where the client can show forms (MCP elicitation), play_in_forms lets Studium ask the questions itself. You read a screen or pick a choice in a form, the verdict opens the next form, and no model writes anything in between; the conversation resumes when a step needs your own words or when you close a form to ask something. When the lesson has a concept map, every scene form offers map: the map form lists each concept with its id, returns to the untouched scene when accepted empty, or takes contest_concept and claim (known or unsure) to ask the check questions and then resume that scene. Only a hint, a question or a proposal still needs the conversation, as the help line says. A form never asks you to say "continue", since nothing can be said in it. Ask your assistant to "play the next steps in forms". How clients draw a long screen, a formula, or this map form has not yet been observed in a real client.

With no model in between, the server's own part is small: an answer given in a form or on a panel gets its verdict in under 20 ms at the 95th centile, transport included (same script, --mode forms and --mode panel, three runs of 32 to 40 answers each, the learner's state saved to disk, 6 November 2026). The time a client then takes to draw the screen is not in that figure and has not been measured.

View the source of this page on GitHub