Opens in a new window
The Rogue and Hack ports were done by an individual agent working largely autonomously over a few hours-long sessions. For NetHack I have had a swarm of agents running on a server for nearly two months, both Claude and Codex. I have been spending substantial effort managing them, and the end is not yet in sight. Early on I tried the same hands-off approach that worked for Rogue. The agents would make progress for a while, then get stuck on a bug and spend twenty minutes poking at random hypotheses, each guess requiring a full test cycle. I would come back to find hundreds of lines of speculative changes and no forward motion. So I started building infrastructure. I wrote an AGENTS.md file defining how each agent should work: what to do when a test fails, how to avoid clobbering another agent’s changes, when to stop and ask for help. I codified eight debugging workflows into reusable skill protocols. I directed agents to build a custom diagnostic tool called dbgmapdump that captures the full game state — map, monsters, objects, player status — in a single dump, so an agent does not have to probe variables one at a time. I advised them to build event logs that record hidden state changes as they happen, so that when a bug manifests at step 50 but was caused at step 30, the step-30 anomaly is right there in the log.
，这一点在whatsapp网页版中也有详细论述
_tool_c89cc_emit "48 85 C0" # test rax, rax
面对2026年底近八亿元本息兑付的刚性现金支出，对于2025年全年净利润仅为7.31亿元的新乳业而言，这无疑构成了巨大的资金压力。