L'Era dei Subagenti: Quando un Agente Non Basta Più
Until six months ago, a coding agent was a single process: a model that read your prompt, iterated on the code, and returned a result. It worked. But real repositories aren't single problems — they're ecosystems. And an ecosystem requires an orchestra, not a soloist.
Claude Opus 4.8, released on June 1st, broke the soloist paradigm. With Dynamic Workflows in Claude Code, you can spawn up to 1000 parallel subagents in a single session. Each subagent is an independent instance with its own context, its own tools, and its own reasoning loop. The orchestrator model doesn't execute — it coordinates. And it does so across three effort levels: high (default), extra, and max.
The Architectural Shift Many Underestimate
The most important novelty in Opus 4.8 isn't the number of subagents. It's the reformulated Messages API: it now accepts system entries directly within the messages array. This means you can change system instructions mid-task without invalidating the prompt cache. In practical terms: an orchestrator agent can rewrite the rules for its subprocesses while they're running, adapting strategy in real time without the cost penalty that previously would have made the operation prohibitive.
Translated: you're no longer deploying an army with fixed orders. You're commanding an army whose rules of engagement you can change mid-battle.
The Multimodal Alternative: Qwen3.7-Plus
While Anthropic goes horizontal (more agents, same type), Alibaba has gone vertical with Qwen3.7-Plus: a single multimodal agent that integrates vision (images + video), deep reasoning, tool invocation, and autonomous iteration within the same instance. It doesn't spawn subagents — it does more things itself.
Qwen3 Coder Next, the open-weight variant optimized for coding agents, offers 1M tokens of context at $0.65/1M input. For local workflows where network latency between subagents would be a bottleneck, this "all-in-one" approach has a real advantage: zero coordination overhead.
Which to Use? It Depends on the Problem's Topology
The answer isn't "one or the other." It's: what shape does your problem have?
- Hierarchically structured problems (monorepos with independent modules, multi-step CI/CD pipelines, codebase scans across 100+ files) → parallel subagents. Claude Opus 4.8.
- Dense-structure problems (a single complex file requiring visual reasoning on diagrams, debugging where looking at the error screenshot is faster than reading the log) → single multimodal agent. Qwen3.7-Plus.
In practice, a Watchdog can use subagents to monitor independent microservices (each with its own scope), while the deep reasoning instance for architectural decisions is single and high-context. It's not ideology — it's problem geometry.
Practical Insight: Start with One Subagent, Not a Thousand
Before reshaping your architecture with `spawn_agent` calls, ask yourself this: does your subagent need to communicate with the others to complete the task? If the answer is yes, the coordination cost quickly exceeds the parallelism benefit. Start with 2-3 subagents with disjoint scope, measure real throughput, and scale from there. Claude Opus 4.8 allows this — the default is "high effort", not "max". Use the scale with conscious intent.
Subagents aren't hype — they're an architectural paradigm shift. If you're designing your next agentic system and want to understand when to parallelize and when to consolidate, the problem's topology is your compass.