P-010 · 2026 Hackathon entry · Meta Global AI Developer Hackathon

Meta Game Onboarding Optimizer: closed-loop onboarding playtests with Meta Muse

A Meta Global AI Developer Hackathon project (indie developer track, theme Build for Your World, Oct 26–Nov 4, 2026, Cerebral Valley): Muse agents play a free-to-play browser game, find where new players get stuck, propose a fix, and verify it only after a human approves. I submitted this project for the event; winners are announced on 9 November 2026.

MuseSAM 3.1Multi-agentPlaytesting

How the loop runs

Playtester agents (Muse Spark)sessions with screenshots, actions and events
Analyst agent (SAM 3.1 + Muse Spark)funnel and ranked friction points
Designer agentproposed config or feature patch
Human approval gatenothing is applied until a person says yes
Experiment agentpaired A/B rerun of old vs new config, then a short report

The game under test is Hearths of the Glen, an original Celtic dawn-age village builder built for this project so every fix can be played, not only described.

The problem

The first minutes of a free-to-play game decide whether a player stays. Finding onboarding friction with human playtests and funnel analysis is slow and expensive, and every fix needs another round of testing. I wanted a loop that playtests, diagnoses, proposes a change, waits for a human, and checks whether the change helped.

Approach

Playtester agents use Muse Spark as a computer-use brain. They look at the screen, click through onboarding, and keep a short working-memory scratchpad. Five literature-grounded personas cover different player types (Rush-Clicker, Story-Guided, Strategist, Tinkerer and Hoarder).

The analyst uses SAM 3.1 to mark UI elements in gameplay screenshots and Muse Spark to reason over the session. It builds an onboarding funnel and ranks friction points with linked evidence, including a distractor-to-goal visual salience note from the SAM masks.

The designer proposes a concrete onboarding config patch, or turns a free-text feature idea into a schema-validated feature plugin (Feature Forge), with an emblem from Muse Image or a procedural stub. No patch is applied or playtested until a human approves it.

The experiment agent then reruns the playtesters on the baseline and the patched config as a paired comparison and writes a short Council Report with a clear verdict and side-by-side replay. Bug hunting produces a diagnosis for one run; Feature Forge produces a verdict for one frozen experiment. Those two paths stay separate on purpose.

Models and credits

Model	Role in the loop
Muse Spark	Default playtester brain; diagnosis and plain-language report summaries
Muse Wonder	Coding help during the build; optional bot brain; drafts feature ideas behind Ask Muse Wonder
SAM 3.1	Segments screenshots and marks real UI targets for bot clicks and the analyst's visual note
Muse Image	Game art and feature emblems
Muse Voice Transcribe	Optional bot ears for Elder voice lines (off by default in the menus)

Sound and voice are separate from Meta: Celtic soundtrack, Elder Kaelen voice lines, bot persona voices and sound effects were generated with ElevenLabs and shipped as audio files.

Credits: Meta models — Muse Spark, SAM 3.1, Muse Wonder, Muse Voice Transcribe, Muse Image. Playtester agents play an original browser game, hunt bugs and A/B-test new features through player personas. Music and voices: ElevenLabs (not a Meta model).

Try the demo

AI Agents Playtest a Game's Onboarding | Meta AI Hackathon

Playable demo coming soon

Limits

Agent playtests are a smoke test, not a sample of real players. Persona findings are hypotheses to check with people. SAM 3.1 marks detected regions; it does not prove what a human would notice. Bot think-aloud is recorded commentary, not a window into hidden model motives. Runtime changes are schema-validated JSON patches only, and every patch stops at the human approval gate. The public demo on Hugging Face uses demo mode without live Meta calls.