“Graph” is a data structure, not a quality guarantee. I connected code-review-graph to IBM Bob over MCP and benchmarked both on Open Liberty: 120,016 tracked files. The graph runs were 18.8% faster and returned 18.7% less tool output—but found 1/10 required call sites. Solo Bob found 10/10. The useful lesson is how to benchmark agent tooling without rewarding fast wrong answers. https://www.the-main-thread.com/p/ibm-bob-code-review-graph-benchmark #IBMBob #MCP #CodeReview #OpenLiberty