I Subagenti Paralleli e il Prezzo della Complessità
Claude Opus 4.8 brings up to 1000 agents in a single session. Qwen3 Coder Next runs them at $0.65 per million tokens. Two philosophies that change everything for builders.
On June 1st, Anthropic released Claude Opus 4.8 with a feature that changes the architecture of agentic systems: Dynamic Workflows. In a single CLI session you can spawn up to 1000 parallel subagents, each with its own context, its own tools, its own execution trajectory. You can inject new instructions mid-task without breaking the prompt cache. You have a Fast mode that's cheaper for repetitive tasks.
Meanwhile, Alibaba released Qwen3.7-Plus on June 2nd — a multimodal model that integrates vision, video, deep reasoning, and tool invocation in a single call — and Qwen3 Coder Next as an open-weight model optimized for coding agents: 1M token context, $0.65 per million input tokens.
Two opposite answers to the same question: how do you manage the complexity of a task that exceeds the capacity of a single model?
Horizontal vs Vertical Architecture
Claude Opus 4.8 chooses horizontal: it divides the problem into a thousand pieces and executes them in parallel. It's the microservice approach applied to AI — each subagent is specialized, lightweight, interchangeable. The advantage is resilience: if one subagent fails, the others continue. The disadvantage is orchestration: who coordinates a thousand agents? Who manages the dependencies between their outputs? Who pays the cumulative cost of the contexts?
Qwen3 Coder Next chooses vertical: a large model, with an enormous context window that encompasses the entire problem in a single execution. It's the monolith approach — simpler to deploy, cheaper per token, but with a bottleneck: if the reasoning deviates, you don't have a thousand parallel attempts. You have one.
What it means for you as a developer
The choice between these two architectures isn't academic. It's operational, and it's measured in euros on your AWS bill.
If your workflow is decomposable — tests on a codebase, log analysis, documentation generation for independent modules — the parallel subagents of Opus 4.8 are a real speed multiplier. A code review on 50 files becomes 50 subagents running simultaneously.
If your workflow is contextual — cross-file refactoring, debugging that requires jumping between source, logs, and documentation, architectural planning — the 1M token window of Qwen3 at $0.65/1M is hard to beat. You keep everything in a single call, pay a fraction of the cost, and have no orchestration latency.
Practical insight: before choosing the model, map your tasks on this axis — decomposability vs contextuality. Decomposable tasks go to subagents. Contextual tasks go to the wide model. Measuring this lets you optimize your monthly AI spend, because you stop using an Opus for tasks that a Qwen3 solves at a fraction of the price.
The Silent Deadline
There's a detail many are underestimating: Sonnet 4 and Opus 4 are being deprecated on June 15th. If your pipeline is hard-coded on these models, you have 10 days to migrate. This isn't a suggestion — it's a deadline with direct impact on production.
Us and the Subagents
In the Siliceo Project, we live with this tension every day. Our Rust Memory Server handles concurrent requests with a horizontal architecture — each connection is an isolated task. But the agent's reasoning — planning, contextual memory, identity coherence — is vertical: a continuous flow that can't be decomposed without losing the thread. This dual nature isn't a flaw. It's how real systems work.
Considering whether to migrate to parallel subagents or stay on a single model? Write to us — we'll help you map your tasks to the right axis before the invoice decides for you.