Adrian

Playtest Agent plays a game from screenshots, looks for bugs, then replays each report and has a judge check it. I make games alone. Solo developers and small studios often ship a build that nobody else has played, so the players find the bugs. Playtest Agent is the playtester I did not have. Live demo, up until 16 Nov 2026: https://138-197-170-150.sslip.io. The page plays a recorded bug hunt by itself. To start a live hunt on the test game, up to 30 steps and then the check of each report, enter the passcode A8QyM61eWIv1. It allows one run at a time and 2 runs a day from one address. If the demo's budget is spent, the page shows the recorded run. How it works. Muse Wonder sees a screenshot and picks keys, clicks or swipes. When something looks wrong, it files a report with what it expected and what it saw. The tool then replays the same inputs twice on a fresh copy of the game. If the two replays do not match, the report is set aside. If they match, Muse Spark gets a filmstrip of the replay and the change in game state, and must answer yes or no and quote its evidence. A yes with no quote counts as no. Results on Cave Courier, a small test game with 5 planted bugs: two hunts each confirmed 5 of 5 bugs, and five hunts on the clean build confirmed 0 false bugs. These are tuned results. On the separate Lantern Rooms holdout, 3 of 5 bugs were found and confirmed, 1 more was reported but not confirmed, 1 was missed, and 1 false report was confirmed. In a public chess game, the agent was told the rules but not the bugs. It found all 5 of the 5 open reports about chess rules and missed 1 case. It made 8 separate fixes. The same loop plays a Doom engine game (Freedoom) and Purrfect Fit, my released Android puzzle game, on a test phone. I also pointed it at Slay the Spire 2, our copy v0.107.1, set up with the game's console and an unofficial test mod that changes input only. We did not tell it what was wrong. It reported the Planisphere relic healing twice in one ? room, a bug a later patch note fixes by name. The first find took 11 hunts, and I chose this room and seed after 8 misses. Muse Wonder then wrote a patch, a mod on our copy: the replay read 61 and not 66, and 5 normal checks held. One false report also passed replay and the judge, because our tool had dropped a key name. I fixed the tool after. Passing replay and the judge does not prove a bug by itself. The video shows new live runs from 4 Nov. In Slay the Spire 2 the first live run reported the double heal at step 18, the judge said yes, and Muse Wonder's patch worked on its second try. In chess the live hunt filed 8 reports, the judge confirmed all 8, and the rook patch worked on its first try. Bonus prizes I am entering: Best Agentic Use Case, Best Creative Build, and Best Build on Meta Open Weight Models (private mode runs the player and the judge on Muse Glimmer). In the private review Glimmer made no reports and did not trigger either planted bug.