๐Ÿ“ฃ Fall 2026 Lyceum Institutional is now open for university enrollment โ€” see the pilot partnership program โ†’
๐Ÿค– Anatomy of the Agent

Plans. Tools. Memory.
Action.

Lyceum isn't a chat box wrapped around an LLM. It's a full agentic system that plans multi-step study paths, calls verified tools, retrieves from your textbooks, and stays with each student across semesters โ€” exactly the way a great human tutor would.

๐Ÿงญ
01 ยท PLANS

Multi-step planning

Given "I'm stuck on series convergence," the Agent doesn't just answer. It diagnoses prerequisite gaps, builds a 4-step path through the right concepts, and schedules a follow-up.

โŒ– diagnose โ†’ ๐Ÿ“ plan path โ†’ ๐ŸŽฏ execute โ†’ ๐Ÿ” schedule retry
๐Ÿ› ๏ธ
02 ยท TOOLS

Real tool use

The Agent calls Python (sympy / numpy) to verify derivations, retrieves from your uploaded textbooks via vector search, runs Wolfram for symbolic math, and queries the curriculum graph โ€” every call is logged and auditable.

๐Ÿ Python sandbox ๐Ÿ“š Textbook RAG ๐Ÿงฎ Wolfram ๐ŸŒ Web search ๐Ÿ“Š Knowledge graph
๐Ÿง 
03 ยท MEMORY

Cross-semester memory

The Agent remembers Maya struggled with LIATE in week 7, that she prefers visual derivations, that she always reviews on Sunday nights โ€” and weaves all of it into how it teaches her in week 14.

๐Ÿ“ episodic ๐Ÿ“‹ semantic ๐ŸŽจ preference โฑ scheduling
โšก
04 ยท ACTION

Autonomous workflows

Overnight, the Agent runs jobs without prompting โ€” generates next week's practice set, drafts a class digest for Dr. Hartman, and pages the dean of students about a struggling cohort it identified at 02:14 AM.

๐Ÿ“„ Practice generation ๐Ÿ“จ Faculty digest โš Advisor alerts
๐Ÿ”
Multi-turn agentic loops, not single-shot prompts. A typical hard tutoring session at Lyceum is a 14-step trace: plan โ†’ retrieve โ†’ reason โ†’ tool-call โ†’ verify โ†’ respond โ†’ wait โ†’ diagnose โ†’ re-plan โ†’ ... The student sees a teacher. Internally, an Agent is running.
๐Ÿงฌ The frontier stack

Six frontier models,
one routing brain.

The Agent picks the right brain for each request โ€” deep reasoning for proofs, fast clarifications for follow-ups, long-context for full-textbook ingestion. Multi-vendor by design โ€” never a single point of failure for your students.

Tokens ยท last 30 days
12.4B
+18% MoM
Frontier models routed
6
across 4 providers
Median first-token
380ms
P95 ยท 1.2 s
Inference uptime
99.97%
90-day rolling
๐Ÿงฌ Active routing โš™๏ธ Failover topology ๐Ÿ“Š Cost per task
โ— All providers healthy 6 models
Aa Model โš™๏ธ Provider โŠž Routed for โŒ— Share โšก P50
๐Ÿง Claude Opus 4.7
Anthropic
Multi-step proofsLong humanitiesAmbiguous tutoring
28% 2.1 s
โšกClaude Sonnet 4.6
Anthropic
Everyday tutoringSocratic dialogueConcept explanation
34% 820 ms
๐ŸŒฌ๏ธClaude Haiku 4.5
Anthropic
Fast clarificationsFollow-upsSmall steps
14% 320 ms
๐Ÿ’ปGPT-5.1
OpenAI
Code generationCS walkthroughsDebugging
11% 1.3 s
๐Ÿ“šGemini 3 Pro
Google
Long-context (whole textbooks)Deep Think ยท exam analysis
9% 1.6 s
๐Ÿฆ™Llama 4 Scout
Self-hosted ยท Meta
High-volume FAQCost-optimized MoEAir-gap deployments
4% 210 ms
๐Ÿ›ก๏ธ
Multi-vendor by default โ€” never a single point of failure. When one provider has an outage, the router auto-fails over to a backup. In Q1 2026 we maintained 99.97% inference uptime through three documented provider incidents.
๐ŸŽ“

Give every student
a teacher at 2 AM.

A 30-minute demo is enough to tell whether Lyceum fits your campus. We'll walk through three of the most common student help-requests at your institution โ€” and show you exactly how Lyceum handles them.

Avg. response time < 4 hours ยท serving institutions across U.S., Canada, and Europe ยท 50+ partner campuses