The Second Agent: Controlled Experiments on Recursive Delegation in Coding Agents

A coding agent can solve a repository task itself or delegate parts of it to other agents. We ask when that second agent is worth its cost. We test two one-shot designs: patch returns and written findings returned to a single writer. Across two experimental campaigns on verifier-scored repository repairs, recursive trees used roughly six times the tokens at seven nodes and twelve times at fifteen nodes, while no primary verified-correctness contrast resolved a gain over one agent. Execution traces show why added computation need not become complementary work: siblings converged on the same files beyond task locality, and 90.3% of replayable refused sibling patches contained genuine line-level conflicts. We then tested a seven-node single-writer design in which workers returned findings instead of patches. The intervention verifiably removed shared-write integration, but its verified-success difference from shared write was +1.25 percentage points (95% CI [−9.76, +12.82]), leaving no resolved gain at this sample size. Finally, exploratory static pre-task routers did not show a reliable gain over the training-chosen constant; one fitted policy selected the single agent on every held-out task. At the tested scales, these two designs purchase more compute without a resolved verified-correctness gain. A remaining research question is whether execution evidence can identify when another agent earns its cost.