Why cross-vendor review is a validation error, not a convention
In bary, must_differ_from isn't a suggestion — a pipeline where an agent reviews its own code never starts.
Most multi-agent coding tools let you add a "review" step. Few of them stop you from pointing that review step at the same agent that wrote the code.
In bary, that's not allowed:
steps:
- name: implement
agent: claude-code
prompt: "Implement: {task}"
- name: review
agents: [claude-code, codex] # claude-code reviewing itself
must_differ_from: implement
prompt: "Review the latest commit."
Running bary validate against this file fails before a single token is spent:
error: step `review`: reviewer `claude-code` is the same agent that ran
`implement`; cross-vendor review requires a different agent
Why this matters
Self-review isn't just a theoretical risk. A single vendor's agent reviewing its own output inherits the same blind spots that produced the bug in the first place — the same training data, the same failure modes, the same confidence in its own reasoning. A second vendor's agent is a genuinely different reviewer, not a rubber stamp with extra latency.
This is also why the check happens at validation time, not run time. If must_differ_from were just documentation, it would be the first constraint teams quietly disable under deadline pressure. Making it a parse error means there's no pipeline to run until the review is real.
What this doesn't solve
Cross-vendor review doesn't guarantee correct review — two agents from different vendors can still both miss the same bug, or get into a rubber-stamp pattern if the same two agents always review each other. That's part of why we're building TeamBench: to measure, rather than assume, which team compositions actually catch more issues.
Try it yourself:
bary init
bary validate # try breaking must_differ_from and see what happens