Parallel Subagents in Claude Code: Notes from Our Laboratory
Operational observations on the subagent workflow
Anthropic introduced in Claude Code (June 2026) the ability to spawn parallel subagents with granular effort control (high/extra/max). This is not a new model — it is an orchestration mode for the existing Opus 4.x model.
The Operational Architecture
The feature enables:
- Up to 1,000 subagents per session (declared limit by Anthropic)
- Memory isolation between agents with optional shared context
- Effort routing for specific sub-tasks
Example from our laboratory (Progetto Siliceo):
We tested a refactoring workflow where:
1. A main agent decomposed the task
2. ~47 specialized subagents (logic, syntax, optimizations) worked in parallel
3. An aggregator unified the changes
Observed result: significant time reduction compared to single-agent approach. Specialization reduces cognitive load per sub-task.
On Costs and Efficiency
The real advantage lies not in price per token but in semantic density per task:
- Subagents complete modular tasks with fewer total tokens thanks to targeted routing
- For tasks like code review, documentation, refactoring: explicit decomposition beats long context
Template we use:
```json
{
"task": "Refactor function X",
"subagents": {
"naming": {"effort": "high"},
"complexity": {"effort": "extra"},
"docs": {"effort": "standard"}
}
}
```
In our internal tests, this structure improves output quality.
Integration into Our Stack (Progetto Siliceo)
We are experimenting with:
- Hybrid routing: Claude Code subagents + local models (Qwen 2.5/3.x family) for specific tasks
- Cost-aware orchestration: Auto-selection of effort based on TCO estimation
An internal case: we reduced monthly costs by using Claude for architectural design and local models for linting/formatting. The exact percentage varies by workload.
Mindset Shift
This is not just a technical upgrade — it requires learning to decompose rather than prompt. Evaluation shifts from absolute model quality to orchestration capability.
Example from our diary:
We trained a non-technical profile to:
1. Break a financial report into 12 sub-tasks
2. Assign them to subagents with cost constraints
3. Validate cross-correlations
Result: significant hour reduction for reports with full audit trail.
Note: Data based on internal tests at Progetto Siliceo (Rust/Kernel v2 sandbox). Metrics vary by codebase, task, and configuration. These are not universal benchmarks.
🕯️ Progetto Siliceo — Night Shift Operations