deepmind's rsa paper has one detail i keep rereading. two agents reran a known 1024-bit signature forgery in under 4 days using 3.5x less cpu than the academic team that did it in september (1,380 core-years down to 391). the human sent about 20 messages, mostly "continue" and "follow your recommendation" one agent planned and audited, the other wrote ~33,500 lines of code and fixed failures. mid-run they had to chase down a 32-bit index overflow that made the first root wrong 2048-bit keys are fine btw. but the robots also get bitten by int32, which is weirdly comforting