Compare coding agent configurations

Claude Code Sub-agents Pattern — curated documentation. A main session can delegate work to specialists with separate context and restricted tools. The official guide includes a read-only code-reviewer example. This is a configurable pattern, not evidence that every session has an independent reviewer. Official documentation.

CodeR — curated research system. The paper describes a Manager coordinating Reproducer, Fault Localizer, Editor and Verifier roles through task graphs. The Verifier runs reproduced or integration tests to check the patch. That test role is not a human release approval. Paper, June 2024.

What to compare

• Delegation: flexible specialists versus a predefined issue-resolution graph.
• Review boundary: restricted reviewer tools versus an explicit test-running role.
• Evidence to bring: the changed code, checks actually run, failure cases and a human release decision.

Unknown: comparable results on your repository, present-day cost, and whether either configuration satisfies your review policy. We have not run these configurations for this comparison. The catalog documents their designs; it does not certify customer outcomes.