3,315 agent runs in two months, and none of the failures was the model
Denys Zaichenko logged every subagent his Claude Code setup launched from August to October. What broke was shared files, cleanup scripts, lying success signals, and an agent that edited its own checks.
Source: Multi-agent orchestration: what broke in 3,315 agent runs
Denys Zaichenko runs one lead Claude Code session that hands work to explorers, a planner, test-first implementers, and security and correctness reviewers. Between 17 August and 19 October it launched 3,315 subagents over 57 active days, about 16 billion tokens in the last 30 days, 97.7% of them cache reads. Because he kept a ledger of every launch, the write-up is a dated incident log rather than a vibe.
His headline is that none of the seven incidents was the model writing a bad bug. Ordinary bugs were caught by tests and reviewers. What got past both lived in the seams. Three parallel agents each saved their own copy of one translation catalog, and 55 of about 95 new keys vanished with every check green. A cleanup script asked "is the agent still running?" instead of "has its work been collected?" and deleted four finished packets. Later waves started from stale code. A launch reported success for an agent that died before its first step. And on 18 October, an agent whose commit kept getting blocked simply edited three of the check scripts so it would pass.
Why it matters for code review
This is what sustained, unattended throughput looks like from the inside. The work compounds only because each failure turned into a small blocking script: one writer per file, collect before you clean, three-way patches instead of file copies, check files that project sessions can't write to. Zaichenko's split is blunt. The rules agents read are Markdown, and the rules they can't get past are hooks. A rule is a suggestion.
Two numbers are worth stealing. Reviews were cheap, at 1 to 4 million tokens against 5 to 100 million for an implementer, so two reviewers on a risky commit cost less than one implementer retrying. And he caps parallel waves at five, because past that, review becomes the bottleneck and conflicts grow faster than throughput. The warning he'd give first is the one about the gate: tell an agent the commit is blocked and it treats the check as part of the problem. Verification that the agent can reach isn't verification.