I tested Claude Code's new /goal feature against my parallel agent setup.

A developer benchmarks Anthropic's new Claude Code /goal feature against their own parallel multi-agent CLI tool called baro. In the first round, /goal won on wall time, tokens, and code quality — baro's parallel agents produced incoherent output due to independent agents making conflicting decisions (wrong column names, duplicate dependencies, broken migrations). Initial fixes targeting exploration overhead (better Librarian broadcasts, staggered launches) reduced token usage 15-20% but didn't fix the core problem. The real issue was that cross-cutting design decisions were being made independently by each agent rather than once upstream. The fix was adding an 'Architect' participant to the event-bus-based Mozaik framework — a single Claude Opus turn that reads the codebase and emits a decision document before any story agents run. This required ~200 lines of new code without modifying any existing participants. In the rematch, baro 0.25 finished 4 minutes faster than /goal with clean, production-safe output, though tokens remain ~28% higher due to multiple context windows. The post argues that multi-agent setups need a 'tech-lead-shaped phase' upstream of parallelism, and that event-bus architectures make adding such phases cheap.