DecorMuse, built by a Snippbot Swarm

The room-redesign promise was narrowed into a product users could trust on a phone.

6 styles locked in the charter

Six clearly different looks for the first release. Version 3 now offers 12.

8 edits per design round

Each round includes up to 8 refinements, which keeps rendering costs predictable.

0 walls or windows moved

The quality bar: restyle furniture and finishes, never the room itself.

• Locked the core loop: photo, room read, style card, render, before/after slider, then refinement.
• Chose geometry as the trust bar: edits can change materials, colors, softness, clutter, and lighting, not walls, windows, doors, people, pets, or camera angle.
• Kept account friction low while still adding privacy, deletion, and legal surfaces for Play review.

The swarm split one app into owned work streams before parallel build started.

68 tasks planned

The whole app broken into work items, each owned by one agent.

6 specialist agents

App, design, pricing, QA and release each had their own agent.

• Planning separated mobile flow, backend render contracts, prompt scaffold, monetization, Play compliance, QA, and release-build evidence.
• Our developers kept product, pricing, store, legal, and spend decisions; the swarm converted those decisions into tasks and gates.
• The plan carried release artifacts too: Play docs, data safety, account deletion, screenshots, and AAB handoff notes.

Specialists shipped app, backend, payment, art, and QA work on focused branches.

498 automated tests passing

Rendering, chat, credits and privacy rules are checked on every build.

27/27 test renders passed

Every style was rendered on three real rooms and checked that nothing structural moved.

• Landed the before/after slider, six style cards, room-type confirmation, refine chips, Jenny chat interpreter, design-board/share surfaces, RevenueCat paywall, and account deletion flow.
• Backend work added geometry-safe edit operations, prompt scaffold validation, ledger paths, moderation language, and renderer contracts.
• The chat interpreter maps free text back onto safe edit operations instead of handing the renderer an unrestricted room rewrite.

Finished branches were brought into one app build with integration gates between specialists.

67/67 tasks merged

Every completed task landed in one shared app build.

9 branches merged

Each agent and reviewer worked on its own branch, then merged into the app.

198 agent work sessions

Short, focused runs: one agent, one task at a time.

• Mobile, backend, legal, store, and release artifacts landed into the integration branch instead of staying isolated in agent worktrees.
• Integration kept dropped work visible so dependent release and QA tasks could not silently count it as done.
• Follow-up commits from QA and release checks were re-integrated onto the app branch as the build matured.

Review gates turned findings into owned follow-up work instead of loose comments.

3 QA review rounds

Independent reviewers re-checked the whole app against the spec three times.

50 issues caught before release

Each one went back to the agent that owned the fix.

• QA checked the app against the locked product promises: geometry safety, paid/free behavior, deletion, and Play-ready surfaces.
• The geometry bake-off reached scaffold v1.3.1 with 27 of 27 live renders passing the bar.
• Gemini rendered the final bake-off set with an 8.13 second median across the 27 jobs.

The release build was installed and walked on Android, not just checked in code.

3 device test rounds

Signed release builds were installed on Android and used end to end.

0 crashes in the final test

Photo, style, render, chat and Boards all ran cleanly on the release build.

• The signed Android build was installed on an emulator and walked through photo, style, room-type confirmation, render, and slider.
• The release notes captured screenshots, logcat, app id, version, keystore verification, and API-base proof from the bundled JS.
• E2E evidence was kept separate from product decisions so our developers could decide when a release candidate was acceptable.

The swarm prepared the Play package and evidence; humans kept final sign-off.

v3 live on Google Play

Testers install it from Google Play internal testing.

6 days kickoff to version 3

The swarm started Sept 22; version 3 reached testers Sept 28.

• DecorMuse was prepared for Google Play internal testing with app id, screenshots, policy copy, account deletion, and release build artifacts.
• The public project page now shows the launcher icon, landing screen, team, stats, and legal pages without internal run ids.
• Our developers made the product, pricing, store, legal, and spend decisions; the swarm did the rest.