HAMZA
SAJID
Start a conversation

Multi-agent learning

Omni-Tutor AI

A routed learning system that holds learner context and selects an appropriate teaching mode for the task.

02 / System briefOpenAI · Next.js · Python
Omni-Tutor AI project previewOperational system / 02

01The opportunity

Start with the work that is painful enough to change.

Learning tools often respond to prompts but do not adapt their method to the learner's current understanding, goal, or feedback.

Primary userIndependent learner
System ownerLearning product or educator
Decision-makerProgram or product owner
Frequency

Repeated throughout a learning session and whenever a learner gets stuck.

Friction

A generic answer can explain a topic yet miss the learner's level, misconception, or preferred path to practice.

Cost of the status quo

Lost momentum, repeated questions, and weak signal about whether real understanding improved.

02The work, redesigned

From a fragmented handoff to a controlled operating path.

Before

  1. 01Learner question
  2. 02One generic answer
  3. 03Learner retries
  4. 04Context is lost
  5. 05More confusion

After

  1. 01Learner intent
  2. 02Profile + lesson context
  3. 03Teaching-mode router
  4. 04Focused interaction
  5. 05Feedback loop

03The specification

Make the job, constraints, and definition of done explicit.

Job to be done

Guide a learner through the next useful step while preserving enough context to adapt the following interaction.

Inputs
  • Learner question
  • Current goal
  • Prior attempts and feedback
  • Available lesson material
Output

A teaching response matched to the task, plus a lightweight signal for the next learning turn.

Done means

The learner receives an appropriate mode of help, the response stays within the available learning material, and the next state is captured clearly.

Why AI belongs here

The system must interpret open-ended questions, decide whether to explain, quiz, scaffold, or review, and adapt the next step from learner feedback.

What stays deterministic

Progress records, mastery thresholds, content permissions, and assessment scoring should use explicit application rules rather than model judgment.

04The operating system

A worker is more than a model response.

01Learner goal
02Learning context
03Router
04Teaching specialist
05Feedback capture
06Next-step state
Worker responsibility

Use learner state and task type to select a teaching approach, produce a focused interaction, and record the signal needed for the next turn.

Worker actions
  • Interpret the learning goal
  • Select explanation, practice, or review mode
  • Ground the response in lesson context
  • Capture feedback for the next turn
Always available context
  • Current learning goal
  • Learner level
  • Safety and teaching policy
Retrieved only when needed
  • Relevant lesson material
  • Prior attempts
  • Assessment and feedback signals
ToolPurposePermissionBoundary
Lesson retriever

Find relevant approved material

Read-only

Course-scoped retrieval

Teaching-mode router

Choose the response strategy

Internal

Routes only among defined modes

Progress store

Persist the next-step signal

Scoped write

Learner state only

05Controlled autonomy

Give the system room to reason. Keep the consequences governed.

01

Deliberate routing

Multiple roles exist because teaching mode selection and teaching itself are separate jobs; they are not agents for novelty.

02

Grounded instruction

The response is constrained by relevant lesson context instead of a single all-purpose prompt.

03

Explicit learner state

Progress is stored as application state so the system does not rely on a fragile chat history alone.

06Proof and iteration

Make the system inspectable before asking people to trust it.

What the operator can observe
  • Selected teaching mode
  • Context retrieved
  • Learner feedback
  • State transition
  • Escalation or retry
What the system should be evaluated on
  • Mode selection quality
  • Grounding in lesson material
  • Assessment alignment
  • Helpful next step
  • Safe fallback behavior
Engineering lesson

A more capable answer is not always a better lesson. The system needs to know when to ask, demonstrate, test, or slow down.

Next iteration

Build an evaluation set around common misconceptions and use those outcomes to improve routing rules and reusable teaching skills.

Intended outcome

Instruction that responds to the learner, not just the prompt.

This system briedev.hamzasajid@gmail.comonal design: the work to improve, the boundaries that keep it safe, and the evidence needed to operate it responsibly. It does not claim unmeasured benchmarks.

Discuss a similar system