Quick start

How to use Agentic Lesson Studio

A human-governed studio where instructional designers and specialised AI agents co-author lessons, and instructors review every learner-facing AI response before it is delivered. Nothing an agent produces reaches a learner without an approved human decision.

For instructional designers

Author lessons through the 12-step workflow in /studio.

1. Draft the specification. Open /studio, create a project, and fill in audience, outcomes, constraints and misconceptions. Use Spec Assistant for AI suggestions — you decide what to keep. Approve to unlock the agents.
2. Run each agent in order. Analysis → Learning Design → Subject → Assessment → Content → Learner Simulation → Quality → Accessibility. Each step is locked until its prerequisites are approved.
3. Review, edit, approve. The review page shows the AI draft in a structured editor with model, tokens and cost. Options: Approve, Edit & approve (marks provenance as human), Rerun with instructions (guide the agent), or Request revision(cascades downstream steps to stale).
4. Check the canvas. /studio/<id>/canvas shows every artefact in one read-only view for a final sanity pass.
5. Publish. When all 12 steps are approved with no stale artefacts, publish creates an immutable lesson_versions snapshot. Learners only ever see published snapshots.

Open studio Model settings

For runtime instructors

Review AI output learners see, from /instructor/queue.

1. What arrives in the queue. Two kinds of items: Tutor messages (learner clicked “Ask for help”) and Assessment judgements (AI-scored open-ended responses). Both are hidden from the learner until you act.
2. Read the context. Each row shows the learner’s message or response, the AI draft, and any cited sources. The queue auto-refreshes every 6 seconds.
3. Choose an action.
- Approve — release the AI draft to the learner as-is.
- Edit — rewrite, then Save and approve. Provenance switches to human.
- Regenerate — discard and ask the agent for a new draft.
- Reject — block the response; the learner is told help is unavailable.
4. Comments are logged. Every decision (with optional comments) is recorded against the run for audit and evaluation.
5. Escalate uncertainty. If a learner’s question is out of scope or a judgement looks unsafe, reject with a comment rather than approving a marginal draft.

Open queue

For learners

Work through a published lesson at /learn/<lessonVersionId>.

1. Sign in and open the lesson. Use the link your instructor shared. You’ll only ever see lessons that have been published and approved.
2. Move through the 8 stages. Welcome → Diagnostic → Worked example → Guided practice → Independent task → Reflection → Summary → Delayed retrieval. Progress is saved as you go.
3. Answer honestly. Diagnostics and open-ended tasks help the studio adapt. Open-ended answers are scored by an AI assessor and then checked by a human instructor before you see feedback.
4. Ask for help. Click Ask for help when you’re stuck. The tutor reply is drafted by AI, reviewed by an instructor, then delivered — so short waits are normal.
5. What you can trust. Every lesson snapshot is immutable and human-approved. Anything an AI writes to you at runtime has been checked by a person first.

Tip: If a help request seems to hang, it’s waiting on instructor review — the response will appear automatically once approved.

Governance at a glance

Provenance badges tell you whether an artefact came from an agent (ai), a human edit (human), or the system (application).
Append-only history: AI originals are preserved even when a human edits or overrides them.
Locked steps can’t be started until prerequisites are approved — the workflow enforces order deterministically, without an LLM orchestrator.
Learners never see AI output that isn’t either part of a published snapshot or explicitly approved by an instructor.

What Agentic Lesson Studio is

Agentic Lesson Studio is a live AI-assisted prototype that demonstrates how a human instructional designer and a group of specialised AI agents collaborate to analyse, design, develop, test, publish, deliver and evaluate a lesson. It is not a general autonomous agent platform: it is a controlled demonstration of a complete instructional-design lifecycle where the human remains accountable for every decision.

The principal message is that agentic instructional design is not merely automated content generation. It is the coordinated execution of a complete instructional-design lifecycle under human-authored specifications, explicit quality controls and accountable approval gates.

The closed-loop lifecycle

The application demonstrates this progression from intent to improvement:

1. Educational intent
2. Analysis
3. Learning design
4. Subject research
5. Assessment design
6. Content production
7. Learner simulation
8. Quality review
9. Accessibility review
10. Human publication approval
11. Learner delivery
12. Human-approved tutoring
13. Learning evaluation
14. Proposed revision

Non-negotiable governance rules

Every named agent makes a live server-side OpenRouter request.
No agent output is marked approved automatically.
No workflow step advances until its required output has been approved by a human.
Human approval is a separate database event, not part of an AI completion.
Agents may recommend approval, rejection or revision, but cannot enact those decisions.
The deterministic workflow controller decides which steps are available based on recorded approval states.
A human may approve, edit and approve, request revision, reject, rerun or compare with a previous run.
When a human edits an AI artefact, the original AI output is preserved and a separate human-edited version is created and recorded.
Runtime tutor messages are never sent directly to learners; they enter an instructor approval queue.
Private model reasoning is never exposed; only final structured output, decision basis, assumptions, evidence references, risks, confidence and usage metadata are stored.
All model calls occur through backend server functions.
The OpenRouter API key is never placed in frontend code, browser requests, application logs or public database fields.

Roles

Instructional designer: Creates and edits the lesson specification; runs agents; inspects prompts and inputs; compares agent runs; approves, edits, requests revisions or rejects outputs; chooses revision targets; publishes approved lesson versions; reviews tutor interventions and evaluation recommendations.
Instructor or reviewer: Reviews approved lesson content; approves or rejects runtime tutor interventions; reviews learner escalations; approves AI-generated assessment judgements; inspects misconception patterns. In this prototype the designer and instructor may be the same authenticated user.
Learner: Completes the diagnostic; studies approved content; inspects worked examples; completes guided practice; requests help; completes the independent performance task; receives only approved tutor messages; completes the delayed retrieval task.
Administrator: Configures model assignments; enables or disables models; sets token limits and fallback chains; views API errors and estimated costs; manages demonstration data and resets the demonstration project.

Content types and visual cues

The studio distinguishes content visually so you can always see where an artefact originated:

AI-generated — produced by an agent.
Human-authored — written directly by a person.
Human-edited — originally AI-generated, then changed by a person.
Application-generated — produced by deterministic system logic.
Synthetic demonstration data — preloaded for exploration.

Never describe deterministic application logic as an AI agent, and never imply that an output has been reviewed until a human approval record exists.

Need more depth? Open a step’s review page to see per-agent guidance, or ask the Spec Assistant on the specification page for outcome, constraint or misconception suggestions.