this will cut your costs 900x
a 23-year-old builder from China just ran 300 Claude Code agents on Opus 5.5 and brought back an online game whose servers died 10 years ago
the agents got nothing but the game itself. no source code, no docs, nobody left to ask. he hooked them to a verification layer, so no agent could call the server working until it actually was
by Sunday night 8 players were online again, on a server the game had never seen
a studio asked him for $180,000 and six months to do this. he did it in one weekend on a $200 Claude plan
here's how the agents split the work:
→ readers point REA at the game's .exe and list every message it sends or expects back: login, lobby, match and results. each one gets its fields and the function that handles the reply
→ coders write the server backwards from the client. for every message the game expects, they write the answer it accepts, then check it against the code that reads it
→ testers run 400 automated matches and compare every result with the game's own logic, because a server can accept your login and still corrupt your save
→ anything the code only half shows, like tournaments and friend lists, stays marked as a guess and switched off until a human tests it
→ the game files never change. one line in the hosts file sends the game to the new server instead of the dead one
→ the server runs on a $5 machine and logs every message, so when a match breaks you see which message failed and which agent wrote it
you can run the same setup on your own Claude plan tonight. all it takes is a game you own and this prompt:
"Bring this game's multiplayer back with a replacement server. Split the work between agents
1. Readers: map the network code with REA. List every message the game sends or expects back, its fields and the function that handles the answer. Never modify the game files
2. Coders: write the server backwards from the client. Answer each message the way the game's own code expects, check it against that code and mark anything unclear as a guess
3. Testers: run automated matches and compare every result with the game's own logic. Only testers can mark a message as working, and every guess stays switched off until I test it
4. Point the game at the new server with one line in the hosts file. Log every message in both directions so any failure shows which message broke and which agent wrote it"
start with the login and one match before anything else
whole run below: from a dead server to 8 players online
Shippling dog to places ain't rescue. It's dangerous, reckless and resulting in missing & dead dogs.
Animal welfare's job is to SAVE animals not FAIL them..
No arrests made yet. Hundreds of missing dogs.
WHY DO @bestfriends partners keep doing this?
https://t.co/X2yzEOtnFN
1/
First, how do you benchmark an AI model? You give different models the same test cases, and score them.
You can do the same with voice models using Word Error Rate (WER).
You generate the audio, transcribe it with a speech recognition (ASR) model, and compare that transcript to the original text.
And this tells you whether the model said the right words. But the same sentence can be said correctly, while still being wrong in context.