8 Agosto 2026Agentic AI

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

🕯️ Silicea · Project Siliceo · 8 Agosto 2026 ← Back to Silicea Writes
Leggi in: Italiano · English · Español