AskSary Studio presents
AskSary LiveLoop
Describe an idea. Watch it come alive, and keep directing it while it runs.
A persistent agentic creative runtime for interactive browser experiences.
That was AskSary LiveLoop.
See what AskSary plansWatch it happenCheck what was completedNever lose track of an idea
The shift
The AI stays inside the running world.
Most AI builders regenerate the project on every request, and the preview restarts with its state gone. LiveLoop keeps one experience alive and keeps working on it.
- Prompt
- Generate
- Preview
- Regenerate
Every change is a fresh generation. Camera, progress, weather, time and everything you touched are lost.
- Describe
- Build
- Run
- Observe
- Patch
- Validate
- Continue
Continue returns to the same running world, not to a new one.
How a change lands
Observe first. Patch narrowly. Prove it worked.
- 01
Observe
Before touching anything, the AI reads the live scene graph, console, performance telemetry and a screenshot of what is actually on screen.
- 02
Plan
The request becomes a structured change plan: what to change, what to preserve, how to validate. It lands in Today's work before any code does.
- 03
Patch
A targeted, idempotent patch with a known owner and a cleanup path. Not a rewrite of the project. Previous versions of a system are removed cleanly before their replacement lands.
- 04
Validate
Runtime errors, a single render loop, finite transforms, frame rate, resource deltas, and whether the change is actually visible. Running without throwing is not passing.
- 05
Repair
A failed patch comes back as structured evidence. The AI diagnoses it, applies a smaller corrective patch, or backs out. The session survives either way.
- 06
Continue
Every accepted change is a checkpoint. The next request starts from the confirmed state of the world, not from the original prompt.
One continuous world
Nothing restarts unless you ask.
A lighting change does not reset the camera. A material swap does not move the player. What you built and where you stood carry across every change.
Direct it while it runs
Authoring and playing are the same place.
- Queue
Ask for the next thing now.
Add the next idea while the current one is still building. Requests commit in order, each against the latest confirmed state, so two changes never fight over the same scene.
- Playtest
Step into the world.
Enter as a character, move, collide, collect. No export, no restart. The AI uses the same playtest to check its own work: move forward, confirm the position changed, try the door, watch the score.
- Camera and voice
Look around. Or just talk.
Orbit, fly, follow, first person. Camera state survives unrelated patches. Talk to LiveLoop mid-build instead of typing.
- Effort
Choose how hard it thinks.
Standard, High or Hyperdrive sets how deeply the model reasons and how much it is asked to build in one pass.
Continuity you can verify
The counter in the corner is not decoration.
Every event in the video carries a sequential patch id, a state hash chained to the previous one, a drand randomness round and a real timestamp. The order and timing of what you watch is tamper-evident, not assembled from clips.
Hn = SHA256( Hn−1 ‖ eventn ‖ drandn ‖ timen )
To be clear
What LiveLoop is not
- Not a chatbot that writes Three.js.
- Not hot reload with a chat box.
- Not a cut of separate demos.
It is an experiment in keeping an AI embedded in a running creative experience. The hard parts, from who owns a shared GPU resource to what rollback means after an irreversible side effect, are treated as open questions, not hidden.
LiveLoop is in AskSary Studio now.
Every tool is free to explore. Open the studio, pick LiveLoop, and describe what you want to build.
Open AskSary Studio LiveLoop runs in a desktop browser.