Latency and Cost of Multi-Agent Intelligent Tutoring at Scale

Large language model tutors are easy to build in a notebook and hard to run in a real course. We describe ITAS (Intelligent Teaching Assistant System), a multi-agent tutoring system that a graduate quantum computing course used for a semester at Old Dominion University. The system has three layers. The teaching layer is a Spoke-and-Wheel of three parallel specialist agents (Video, Code, Guidance) followed by a Synthesizer, plus a separate autograder that evaluates both the correctness and the approach of checkpoint submissions. The operational layer is four Cloud Run microservices with session state in Cloud SQL and interaction events streamed through Pub/Sub to BigQuery. The feedback layer is a narrow-scope conversational agent that answers instructor questions over per-lesson pseudonymized event streams, addressing what we call the Blind Instructor Problem: LLM tutors accumulate more data about students than the instructor can reach through routine channels. The architecture is a direct response to specific failures of an earlier prototype, and we describe which of those fixes carried forward and which were dropped for this iteration. We report on a pilot deployment (five students, one course, one semester) interpreted as system-behavior evidence rather than learning-outcome evidence: the teaching layer handled 334 chat turns without the task-boundary hallucinations that domain consolidation would have risked, the operational layer captured 10,628 events across five modules, and the feedback layer surfaced two findings the instructor acted on mid-semester. We do not claim the pilot generalizes. We do claim that the system as described is one workable answer to the question of what an LLM-based ITS needs to look like end-to-end to run in a real course.

Multi-agent tool calling is becoming the dominant interaction pattern for LLM-based systems, yet existing inference frameworks treat each tool call as an independent request, re-processing the entire conversation from scratch even though 85-95% of the prompt is unchanged from the previous turn. We present a stateful inference architecture that converts the $O(n_t)$ per-turn cost of conventional serving into an $O(Δ_t)$ delta-only cost: a persistent KV cache lives across turns and advances by ingesting only the new tokens, while a radix prefix cache extends this across interleaved multi-agent traffic and a prompt-lookup speculative decoder accelerates structured output. Against vLLM and SGLang on novel, fully-generated workloads, the reference implementation is $2.1\times$ faster per turn on a 6-turn agentic workflow and $4.2\times$ on the median turn of a 35-turn one, halving end-to-end wall time. The advantage comes from stateful reuse and speculation, not caching.

Every node in a multi-agent large language model (LLM) workflow retrieves context from memory and injects it into its prompt, where those injected tokens are billed as input tokens at the same per-token price as the system prompt and the user query. Production observability tools report total token cost but do not separate the tokens a node generates from the tokens it is handed, so this component of the bill is invisible to the teams paying it. We introduce the Total Cost of Agency (TCA), a decomposition of multi-agent workflow cost into base prompt, inference, memory injection, miss penalty and context-accumulation components, and an exact attribution method: a two-pass, non-billable token count that measures injected tokens directly rather than estimating them from word-count proxies. On a 200-task enterprise benchmark executed against real model APIs, memory injection accounts for 13.6 percent of the variable cost a compile-time optimizer can act on, about 12 percent of the full billed cost, and its share rises from a structural zero at workflow depth one to 27.6 percent at depth six. Injected tokens grow linearly with depth over the measured range (R^2 = 0.9974, depths two through six); a quadratic fit yields a negative leading coefficient, so the data do not exhibit convex growth at these depths. We show the component is controllable at fixed model tier: reducing the retrieval window capacity from 32 to 2 entries lowers injected tokens by 28.7 percent with an accuracy change within seed-level variation. We report in full that our graph-rewriting transforms are approximately cost-neutral in isolation, that two of the five decomposition terms are zero by construction in this harness, and that total workflow cost is dominated by model tier assignment, which we hold fixed and treat as prior work. Prompt caching is not evaluated; all figures are for the uncached case.