What Claude Code’s Multi-Agent Dynamic Workflows Are
Claude Code multi-agent dynamic workflows are a new orchestration mode where Claude writes and runs its own scripts to coordinate multiple parallel subagents, shifting from turn-by-turn interaction to autonomous, pipeline-style execution that delivers only final results back into the context window for the user. Instead of one AI coding agent stepping through a task, dynamic workflows let Claude spin up dedicated subagents that each own a piece of work, run in parallel, and converge on a shared contract. Architecturally, the orchestration logic moves out of the model’s context and into a script that Claude itself generates, which is what makes large-scale parallel subagent runs practical. This changes how developers interact with the system: they describe the outcome, Claude designs the workflow, and the user mostly watches a multi-agent team execute rather than hand-holding a single assistant.
From Single Agent to Five Parallel Subagents
Dynamic workflows became visible to users with Claude Opus 4.8, where Claude Code can now dispatch five or more parallel subagents inside one coordinated run. In Anthropic’s own research preview, a single dynamic workflow session planned a project, defined a shared interface, and launched five parallel agents that each handled a separate component: code complexity analysis, documentation coverage, dependency auditing, test coverage mapping, and final CLI integration. The key difference from a traditional single-agent run is who holds the plan. In older Claude Code sessions, orchestration lived inside the assistant’s ongoing conversation, inflating the context with planning chatter and intermediate steps. With dynamic workflows, Claude writes an orchestration script that calls subagents directly, so the main context receives compact, final artifacts and summary results instead of the entire internal dialog among agents. This separation of concerns is the backbone of the new architecture.
Real-World Performance: Building a Codebase Health CLI
A concrete test shows how these multi-agent workflows perform against a familiar single-agent baseline. In one dynamic workflow run, Claude Code used five parallel subagents to design and implement a codebase-health CLI tool in 6 minutes and 59 seconds. One subagent focused on complexity analysis using Python’s AST module, another on documentation coverage, a third on dependency auditing, a fourth on test coverage mapping, and a fifth on wiring everything into a CLI with argparse and a README. The result included 62 passing tests, two entry points, configurable flags, and a self-generated SAMPLE_REPORT.md as verification. When asked, the five subagents reported a combined usage of 109,237 tokens for the work they performed. While orchestration cost was not directly measured, the test showed that dynamic workflows can plan, build, and self-test a non-trivial tool in under seven minutes without manual step-by-step prompting.
How Multi-Agent Workflows Compare to Single-Agent Sessions
The same tester ran a traditional single-agent session against the identical high-level goal of building a codebase-health tool. The single AI coding agent produced a working JavaScript CLI with a src folder, tests, and a bin/cli.js entry, but the outcome differed from the multi-agent Python implementation that emerged from dynamic workflows. The gap is not only language choice; it is how work is structured. With one assistant, the user remains in the loop, guiding each step and absorbing every intermediate reasoning trace. With dynamic workflows, the user specifies constraints and desired behaviors, then Claude Code multi-agent orchestration handles decomposition, parallel execution, and integration. According to Anthropic’s documentation, moving orchestration instructions into a separate script is what makes larger parallel runs feasible, because the context window no longer carries the burden of all intermediate control messages.
Developers Shift from Coding to Orchestrating AI Teams
For developers, the architectural shift to parallel subagents changes the daily workflow. Boris Cherny, creator of Claude Code, describes moving from prompting a handful of Claudes in parallel to writing loops that prompt and evaluate Claude on their own. In his words, “My job is to write loops.” That mindset echoes the dynamic workflows model, where humans describe the goal and the coordination policy, while Claude designs internal workflows and acts more like a general contractor overseeing a crew than a single handyman. Popular educator Ole Lehmann framed it this way: before dynamic workflows, you supervised one AI helper; now you receive a completed job from an orchestrated team with built-in checks. The result is that coding becomes more about specifying contracts, failure modes, and division of labor for AI coding agents, and less about manually writing or reviewing each line of implementation as it appears.
