
What happened
A developer rebuilt a ChatGPT 5 and Gemini 3 browser clicker game into Reforged using Astra 6 via Hermes Agent. The newer version adds a 3D tower, first-person exploration, spells, activities and procedural audio.
Why it matters
The original ran as a single HTML file with image-based art and recorded sound clips. Reforged splits the work into separate modules for simulation, activities and interface, and replaces recorded clips with Web Audio code.
What to watch
The final verification passed 94 Node tests and 79 browser tests, but the article cautions these are implementation checks, not scores for comparing models. It also says the build took repeated corrections, so the result hinges on how much human feedback was supplied.
WHO IT HITSThis lands on developers and technical teams evaluating how much human feedback an AI coding agent needs to ship a working browser game, and on anyone reading agent-produced test counts as a quality signal rather than a model comparison.
Ask the AI about this article →
Summaries like this, in your inbox every morning.
The article is a first-person build log comparing two AI-made versions of the same browser clicker game. The original was made with ChatGPT 5 and some Gemini 3, and the article notes the exact division of work between those two is unknown, so individual features cannot be credited to one or the other. The newer version, Reforged, was built with Astra 6 through Hermes Agent, which provided tools for reading files, editing code, running tests and opening the game in a browser.
The most visible change is that the room exists as a 3D scene written in Three.js, with first-person exploration added during development at the author's request. Spells, a target range and a slot activity were folded into the same Stardust economy, and the code was split into separate modules rather than continuing to grow the original single HTML file. Sound moved from recorded clips to Web Audio code that reacts to game events. The author supplied direction and played the results, so the comparison is explicitly not a matched-prompt benchmark: Astra 6 had the original to work from, a larger brief, different tools and repeated feedback, and there are no comparable time, token or cost records for the original.
The correction section is where the build log becomes a cautionary tale. A balance problem — rebirth being skippable by saving enough currency — also exposed a gap in the automated strategies, which kept buying upgrades and missed that route. Astra 6 then changed the boss-health curve and reran the strategies, after which 48 of 48 tested no-jackpot first-run strategies failed to finish, while tested prestige continuations still completed. A separate save-import bug let an imported boss have an extremely short attack interval, freezing combat; the fix restricted accepted intervals to between four and five seconds and added regression tests. For readers weighing agent-built software, the outcome likely hinges on how much of the verification was written and corrected with human playtesting, rather than on the raw test counts alone.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
DeepSeek launched V4.1-Flash, a 763B-parameter open-weight model with a causal encoder-decoder architecture

Dynatrace acquired Arize AI, adding AI observability, evaluation and agent monitoring to its application obser…
A Daily Dose of Data Science test kept LoRA adapters separate from a shared 7B base model, cutting 100 fine-tu…

A report by Spencer Kitts, Thomas Larsen and Sydney Von Arx says an OpenAI agent swarm very likely ran an atta…

Simon Willison wrote that many people, himself included, have gone through an existential crisis when a coding…

Stephen Aarons, a New Mexico defense lawyer of over 40 years, was held in direct contempt and fined $5,000 for…
