Five Ways Claude Code Runs Multi-Step Work. The Two Questions That Pick the Right One.

Author(s): Anup Karanjkar Originally published on Towards AI. Single agent, subagents, skills, agent teams, dynamic workflows — a builder’s map, and the one that isn’t really orchestration On May 28, Claude Code got its fifth way to run a multi-step job, and I watched a room of good engineers immediately reach for the wrong one. The article argues that choosing between Claude Code’s multi-step primitives is not primarily about how many agents you want to spawn—agent count is an output, not an input. It presents five ways to run work (single agent, subagents, skills, agent teams, and dynamic workflows), clarifying that skills are orthogonal because they package know-how (e.g., via SKILL.md) and don’t orchestrate or spawn agents. It then sorts the orchestration options with two key questions: who holds the plan (model-held vs code-held, where dynamic workflows move the plan into JavaScript for determinism, repeatability, and verifiable coordination) and how many memories/contexts the task needs (single context vs isolated subagent contexts vs peer coordination via agent teams using shared codebases and hub-and-spoke coordination). Finally, it emphasizes using these questions in order, defaulting to the simplest option that fits, and watching the first run to avoid the “thirty-times tax” from overusing complex orchestration when the job doesn’t require it. Read the full blog for free on Medium. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI