Claude Code dynamic workflows are JavaScript scripts that Claude writes to orchestrate many 📝Claude Code Subagents at once in the background, moving the plan out of Claude's turn-by-turn judgment and into rerunnable code.
A workflow starts when a prompt asks for one in plain words ("use a workflow") or includes the keyword ultracode; /effort ultracode lets Claude plan a workflow for every substantive task in the session. The bundled /deep-research command is itself a workflow that fans out web searches, cross-checks sources, and returns a cited report. Inside a script, agent() spawns one subagent, pipeline() runs one per item in a list, and parallel() runs a set at the same time, while intermediate results stay in script variables instead of Claude's context window.
The /workflows view lists runs, shows each phase's agents and token use, and can pause, stop, restart, or save a run. A saved script in .claude/workflows/ or ~/.claude/workflows/ becomes a slash command of its own. Runs are resumable in the same session and replay completed agents from saved results. The runtime caps a run at 1,000 agents and, by default, 16 concurrent agents; a "Large workflow" warning appears past 25 agents or 1.5 million projected tokens.
Workflows suit jobs that outgrow a handful of subagents or need findings verified against each other, such as a codebase-wide audit, a large migration, or cross-checked research. They are available on all paid plans, need to be switched on in /config on Pro, and draw on plan limits faster than conversational work. 📝Claude Managed Agents offers the same capability through its multiagent configuration.
📝Claude Code Multi-Agent Orchestration compares workflows with subagents, agent teams, and Projects.