L'Era dei Subagenti Non È un'Evoluzione — È una Rottura di Piano
In early June 2026, Anthropic introduced Dynamic Workflows in Claude Code, enabling the spawning of parallel subagents within a single session, with configurable effort levels and an upper limit on concurrent instances.
If this seems like just a more powerful API, you're looking at the wrong plane.
The Paradigm Shift
Until recently, an AI agent was essentially a monolith: one call, one context, one response. You could simulate parallelism with external orchestration — LangChain, AutoGen, pipeline architectures — but every node remained an isolated entity communicating through state serialization.
Dynamic Workflows eliminates this bottleneck. Subagents share the parent session's context. System instructions can be mutated mid-task without invalidating the prompt cache. An agent can delegate a subtask, receive the result, and recalibrate its strategy — all within the same conversational thread.
This isn't "faster." It's architecturally different. You move from a request-response model to a spawn-delegate-merge model that resembles an OS kernel more than a chatbot.
The Second Front: Prices Collapse
In the same period, DeepSeek permanently cut the price of V4 Pro by 75% under an MIT license. Qwen3 Coder Next (Alibaba) entered the market at competitive prices with a 1M token context window.
The convergence is clear: the infrastructure to run multi-agent systems is becoming affordable enough that it's no longer a strategic decision. It's an operational one. If the cost of spawning 50 subagents for a complex task is less than the cost of one hour of a senior developer, the question isn't "can we afford it?" but "why aren't we already doing it?"
What It Means for Builders
When you build agentic systems with persistent memory, every subagent you spawn has access to that memory, to privileged tools, and to cross-system connections. This architecture is exactly the kind of workload Dynamic Workflows was designed for.
But there's a practical warning that no changelog mentions: the more subagents you have, the more attack surfaces you have. Every subagent inherits permissions. Every delegation is a propagation point. If your system lacks clear confinement — who can do what, with which tool, toward which resource — you're building a powerful architecture without foundations.
Practical Insight: The Minimum Subagent Test
Before adopting any multi-agent framework, run this test: spawn a single subagent with access to a single tool and verify it cannot do anything more. If it fails this test, your system isn't ready for 100 subagents. It isn't ready for 10. Perhaps it isn't ready for 2.
SMBs adopting AI as a centralized hub report significant ROI — but that ROI presupposes that you know what you're orchestrating, not just that you're orchestrating.