ParallACT benchmark tests coding agents' coordination in shared codebases

The ParallACT team built 155 paired coding tasks across 11 repositories, testing agents assigned separate requests that must work together. On contradictory tasks, one agent given both requests succeeded 62% of the time, versus 16% for parallel agents using Opus 4.8. The researchers found messaging raised joint success to 32% but used 2.23 times the single agent's tokens; a one-way handoff reached 54%.