Agents are very good at adding things. They rarely take anything out. After four days of building at full speed, on the morning of 23 September I stopped and asked for an audit.
The first audit
I pasted in an audit prompt I had saved. It is read-only, has a coordinating agent give each subsystem a fresh reviewer, and has a line I now think of as the whole point: "Do not force an abstraction. Prefer boring local code when it is already clear."
Then the coordinator audited its own audit: a coverage pass found three concerns that cut across every subsystem (input, collision, the HUD's sliders) and gave them rows of their own.
Under the hood 22 subsystem rows in the end, 25 accepted findings, and one worker result rejected outright: it described door booleans and line numbers that did not exist in the file. The coordinator caught it by re-reading every cited line.
That stuck with me. One reviewer had invented its evidence, confidently, and the only defence was a coordinator that checked.
The decisions only I could make
The audit ended with questions it would not answer for me. I took most in one go: "In terms of the decisions only I can make, the zoom and weather sliders don't need to be on the results screen. The frozen enemy swing is not intended. A castle prop whose model fails to load should not keep its collider." One I was not sure of: "Just explain to me what I'm deciding here." Then: "Ask me some more questions on all this so we can come up with an implementation plan to clean this up."
Two answers grew the scope. Asked about a slider on narrow screens: "Let's add a minimum width to this that you could possibly work on as well, wherever the landscape pixel width is of a semi-modern phone." The game became landscape only on phones. And: "Also, let's remove all those old screenshots that are sitting there. And the old videos." There were 168 of them tracked in git, 131 MB.
Then I left: "subagent driven, let's go. I am AFK so no approval stops or anything, complete the plan. Anything found during implementation, fix if it's a sub bug that was surfaced, add it into the plan, any questions go with recommendations."
Decisions, recorded
A week later, with several cloud threads running, asking me became the bottleneck. One thread stopped and asked whether to start on its rulings. I could not see what it was waiting for: "What are you waiting on me for? If I answer that, and then you can start implementing those new rulings and fixes, but there's some questions you need from me first." It needed three answers, which I gave in one go.
On 30 September the lore threads went the other way: I told them to decide and write each decision down. By that night the record had passed a hundred rows. "What are decisions 80-100?" A line each came back. Then: "So nothing I need to make a decision on?" Nothing. The rows existed so I could overrule one, not so I had to read them all.
Shaky foundations
By 2 October the project was much bigger, with several long-lived worktrees each running for days. I was worried: "there's a large amount of potential duplication or dead code or unperformant work. Everything seems to be taking longer and longer to do simpler things." And: "I want to make sure that we're not working on shaky foundations. I think a really good audit is needed." Then the line that stopped everything: "Let's not start any new work until we've done a full clean up and restructuring the way we work."
This time there were three audits: dead code, drift (one fact stored in two places, a missing guard) and the same simplification prompt. I wanted action on whatever came back: "probably fixing nearly all of them unless there's some that just are not important. Do not wait for me in any of this. I want you to work long and hard to get us to a real clean starting base."
The results surprised me. The dead-code report called the codebase "unusually clean for dead code". The drift audit found the real problem, a Critical one: a guest who started a New game kept the old game's town. The code was not bloated. It was quietly inconsistent wherever one fact lived in two places.
Under the hood Dead code: 15 findings, none above Medium. Drift: 37, one Critical and four High. Simplification: 26 accepted across 42 subsystems. The deletions that did land took out 143 lines with no change in behaviour.
Clean-up as a queue
What changed was how the findings got fixed: not one big clean-up branch but a queue of small ones, each owned by one short-lived agent and named for what it did: save-session, town-truth, one-walk. Each landed with a line in the handoff.
Some needed me, and again I was a step behind: "Not sure on the chest claimed and what shall we do about movement?" Then, a few minutes later: "how do the town people walk? What was the question?" Once it was explained: "Yes go with A after town-truth lands." The town's people now walk by one set of rules.
Under the hood 17 fix branches over two days, from the dead-code deletions to the castle's torches sharing the kit's light pool. The cost was churn: almost every landing touched the handoff, and several commits exist only to keep it under 10 KB.
The process write-up from that day set a rule: audits that fan out across many reviewers run at most monthly. I broke it three days later.
Four independent auditors
On 5 October I ran out of Claude usage and handed the orchestration to Codex mid-stream: "I've run out of credits. I want you to take over." That afternoon I asked for four separate audits: how the game performs in play, how the agents work ("the quicker we can get to the finish here the better, but I want to ensure we don't introduce bugs"), how we make assets, and the code and assets themselves. "I want each of these to be done independently by independent Astra agents, so they can be thorough and report any findings, where you can orchestrate the fixing of any issues." Astra agents are Codex sub-agents. Each got its own brief, never the others' findings, and none could touch product code. Root merged everything into one numbered queue.
The code auditor found that signing out could silently throw away newer work from another profile. The process auditor found a trap in the way I liked to work: because I kept everything on main and pushed as we went, the tool that picked which tests to run compared against main and saw nothing to test. Work pushed to main had hidden itself from verification. That became a warning in the tools.
What I took from it
- An audit is only as good as its coordinator. One reviewer will make things up.
- Ask me for decisions as decisions. Or decide, write it down and let me overrule.
- Fix as a queue. Small named branches, one owner each.
- Agents write little dead code and a lot of drift. The bugs were in facts kept twice.
