“These problems arise when concurrent agents read or modify shared files, artifacts, or environments without proper coordination, resembling data races in traditional concurrent programs. They form the largest main category, accounting for 267 of 804 (33.2%) pattern instances.”

Not a worktree story: the agents here share one output directory, and the study counts what goes wrong. Concurrent writes are 26.5% of all 804 failure instances (62.9% of Codex's), and "Deliverable Overwrite" alone is 14.9%. Example: a Codex sub-agent replaced test5.rs and Cargo.toml with a different module layout, which "invalidates the main agent's prior view of the artifact". The final test still passed, so a green build hides the wasted integration work. Same paper: concurrency costs 1.41–3.31x the tokens of sequential runs and helps mainly on long-horizon tasks. Caveat: these are a main agent's own sub-agents, not independent teams' agents, and the benchmarks are single-repo tasks. Found via a @zhou_xiaoma post (x.com/zhou_xiaoma/status/2112452079826678234).