Skip to content

Cross-Model Subagent Effort Resolution in Claude Code

Cross-model subagents resolve their default effort from their own model, so the parent's session setting does not establish every worker's reasoning level.

Cross-model effort resolution selects a worker's effort independently of the parent's session choice. Claude Code 2.1.317 changed subagents and teammates on another model to use that model's saved effort, or its default (release notes). The rule matters when a delegation changes models without naming effort.

Defaults and overrides

The model configuration reference distinguishes same-model inheritance from cross-model resolution. A worker without its own effort inherits the session level only on the main conversation's model.

Explicit settings still matter. CLAUDE_CODE_EFFORT_LEVEL overrides an invocation's effort parameter, which overrides the agent definition's effort field. The invocation parameter applies to non-fork subagents and persists on resume (subagent effort reference). An effort cap can lower the selected level (model configuration).

The release note adds a model-dependent default; it does not replace this override chain. The separate CLAUDE_CODE_SUBAGENT_EFFORT default shipped in 2.1.316. The cited documentation does not fully place that variable against cross-model saved settings. Avoid inferring an order for their combination (2.1.316 release notes).

Verify a mixed-model run

Use the following review procedure when changing a worker's model:

  1. Record explicit effort settings and caps with the model change.
  2. Inspect /tasks while the worker runs. Its row shows the model and any effort set for that worker (subagent model reference).
  3. Compare the worker's verified result and cost against the previous configuration before adopting the change.

Prefer an explicit agent-level effort when a repeatable worker needs a chosen level. Still check stronger overrides. A saved user preference is a poor experiment specification because the reviewed agent file does not contain it.

This page establishes configuration behavior, not a measured quality or cost gain. Treat model and effort as separate experiment inputs. Keep the task set fixed when testing either input.

Key Takeaways

  • Include resolved effort in the evidence for a model change.
  • Check overrides before attributing a changed worker result to the new default.
  • Test combinations involving the subagent effort default rather than filling gaps in the documented precedence chain.
Emerging — single recent source Last reviewed: 2026-11-06 Applies to: claude-code@2.1.317