Claude Code Gains Five-Level Nested Subagent Support
Key insights
- Claude Code now supports nested subagents up to five levels deep, each running in an isolated context window and returning summaries upward.
- Boris Cherny positioned the five-level cap as a provisional starting point and requested community feedback on whether it fits real workflows.
- Community response was 77.5% positive, with 22.5% skeptical that nesting redistributes rather than resolves token limit constraints.
Why this matters
Context window exhaustion is one of the most common blockers for developers using AI agents on large codebases, and this feature offers a structured hierarchy to distribute that load across isolated subagent instances. Cherny's decision to frame the five-level cap as provisional and solicit feedback signals that Anthropic is treating this as an iterative design problem rather than a finished one. The split reception, 77.5% positive and 22.5% negative, surfaces genuine uncertainty about whether hierarchical isolation actually reduces the underlying problem or reorganizes it.
Summary
Boris Cherny, Claude Code lead at Anthropic, shipped experimental nested subagent support on June 9, 2026, targeting context window exhaustion in complex workflows.
Each subagent runs in its own isolated context and returns only a summary upward, with nesting capped at five levels. Cherny flagged the cap as a starting point, soliciting feedback on real-world fit.
Essentially: (Anthropic, Claude Code) structured hierarchy replaces the flat, single-context approach for large codebases.
- 77.5% positive community reception; 22.5% questioned whether nesting redistributes rather than resolves token limits.
- One developer warned nesting can worsen performance depending on problem complexity.
The five-level ceiling is an explicit experiment, not a final design.
Potential risks and opportunities
Risks
- Developers who build production workflows on the experimental five-level hierarchy may face breaking changes if Anthropic adjusts the cap or its behavior in future releases.
- If nesting redistributes token constraints without eliminating them, teams could invest in hierarchical agent architectures that hit the same scale thresholds as flat agents.
- Poor summary fidelity in deep hierarchies could introduce silent information loss, propagating incomplete context upward without clear error signals for the primary agent.
Opportunities
- Developer tooling vendors building on Claude Code can now design multi-tier agent orchestration patterns that were previously unsupported at the SDK level.
- Teams maintaining large codebases have an early window to prototype subagent delegation strategies and directly influence Anthropic's cap design through Cherny's active feedback request.
- Consultancies and platform teams specializing in AI-assisted development can differentiate by publishing best-practice hierarchical agent templates tuned for large codebase navigation and debugging workflows.
What we don't know yet
- Whether the five-level cap will be raised or lowered based on community feedback, and what metrics Anthropic is using to evaluate the right ceiling.
- How subagent summary quality affects parent agent accuracy across deep hierarchies, given that only summaries are returned upward rather than full context.
- Whether the performance degradation some developers noted follows a predictable pattern by workflow type, and which categories are most at risk.
Originally reported by digg.com
Read the original article →Original headline: Anthropic's Claude Code Lead Boris Cherny Ships Experimental Nested Subagent Support — Five-Level Hierarchy for Context Window Management