E
Egora
OverviewHow It WorksArchitectureDesignDashboard
Teacher sign inOpen dashboard
E
Egora

A teacher workspace for planning courses, reviewing drafts, and publishing approved work to students.

© 2026 Egora

How It WorksArchitectureDashboardTeacher sign in
Architecture Overview

The System Behind
Egora

How Egora keeps course plans, generated drafts, review decisions, and publishing in one connected workflow.

The Triad Architecture

The system separates concerns into three distinct layers, ensuring the curriculum logic is completely independent from any specific AI runtime.

The Brain

Runner-Agnostic Definitions

All agent roles, the state graph, pedagogical knowledge layers, checkpoint definitions, and output schemas. Pure data — no code. Any AI framework can consume these definitions.

The Bridge

Adapter Layer

Translates the Brain's abstract definitions into runner-specific configurations. Swap the LLM provider or execution framework without touching the curriculum logic.

The Engine

Execution Runtime

The runtime that drives the pipeline: loading agent manifests, traversing the state graph, calling LLMs, routing checkpoints, and managing retries.

The Fractal Triad Pattern

Every phase follows the same internal structure. Quality is checked locally at each phase — catching a bad unit plan in Phase 2 is far cheaper than catching it after 20 lessons have been generated from it.

Orchestrator

Decides which workers to activate, routes data to them

Workers (1–8 per phase)

Produce the actual content — each with a single focused role

Evaluator

Validates output against the upstream spec — triggers targeted retries on failure

retry feedback loop

Above the per-phase agents sit two session-persistent agents: the Lead Orchestrator (manages phase transitions) and the Lead Evaluator (checks cross-phase coherence).

The Agent Inventory

37 agents organized into five categories — each with a single job, a single set of context, and the ability to be retried independently.

14

Content Workers

Generate domain-specific content: instructions, assessments, resources, visuals, assembly, QA.

07

Orchestrators

Manage phase execution, transitions, and coordinate which specialists activate.

07

Evaluators

Validate output quality against upstream specs at every phase boundary.

02

Planners

Decompose higher-level specs into lower-level specs (units → modules).

01

Advisor

The Pedagogical Advisor — generates session documents and guards teaching standards.

08

Sub-Agents

Dispatched by parent orchestrators for specialized tasks: video curation, diagram generation, stock photos.

Confidence-Routed Checkpoints

The pipeline has 9 checkpoints classified into three types. Quality checkpoints use the upstream evaluator's confidence score to decide whether to auto-approve or pause for teacher input.

02

Input Checkpoints

Always pauses — gathers data the system doesn't have (scope, audience, pedagogy).

06

Quality Checkpoints

Auto-approves when confidence meets the trust threshold. Pauses with targeted questions when low.

01

Final Checkpoints

Always pauses — non-negotiable professional sign-off before delivery.

How Confidence Routing Works

High confidence — Auto-pass. Teacher never sees it.
Medium confidence — Auto-pass with warnings. Amber badge. Teacher notified.
Low confidence — Pause. Red bell. Teacher gets specific, targeted questions.

Even in Autonomous mode, genuinely bad output still triggers a teacher interrupt. The dial adjusts sensitivity — it never removes the safety net.

Feedback Loops & Adaptive Execution

The pipeline isn't a rigid sequence — it's a state graph with bounded feedback loops that allow targeted correction without cascade regeneration.

Three Types of Back-Edges

QA Retry Loop

If the final QA agent finds issues, the system clears downstream state and re-runs assembly with structured feedback — up to 2 retries.

Checkpoint Rejection

Teachers can reject at any checkpoint. The target agent re-executes with detailed feedback about why it was sent back.

Evaluator Retry

Phase evaluators can trigger targeted retries — "re-run the rubric agent," not "re-run the entire phase."

Conditional Agent Activation

Not every lesson needs every agent. The pipeline adapts automatically based on the pedagogical approach declared in the Course Outline.

Project-Based Learning:

Instructions, Rubric, Critique Protocol, Visual Diagrams

Traditional Direct Instruction:

Instructions, Rubric, Worked Examples, AI Prompts

Same pipeline. Different agent subsets. Static graph, dynamic execution.

Design Principles

The seven rules that govern every architectural decision in the pipeline.

Specs Constrain, Not Prescribe

Each spec defines boundaries and goals. Agents have creative freedom within those boundaries.

Evaluate Locally

Catching a bad unit plan in Phase 2 is cheaper than catching it in Phase 4 after 20 lessons.

Static Graph, Dynamic Execution

The topology never changes. What changes is which agents run, which checkpoints pause, and what questions are asked.

Teacher Controls the Dial

Trust mode lets teachers choose their involvement. Checkpoints interrupt by exception, not obligation.

Agents Are Specialists

Each agent has one job, one handler, one set of context. Specialization enables targeted retry.

Edit Surgically

Leaf edits are free. Spec edits trigger targeted re-evaluation, not cascade regeneration.

Opinionated by Default, Flexible on Override

Non-negotiables enforce baseline pedagogical standards. But the teacher always has the final say.

Built for Technical Evaluators

If you're evaluating AI curriculum platforms for your school or district, we'd love to give you a deep-dive walkthrough.

Open dashboard