From Harness to Dynamic Workflows: Claude Code’s New Multi‑Agent Task Orchestration Paradigm

Thariq, a member of the Anthropic Claude Code team, announced Dynamic Workflows as a major upgrade after skills and sub‑agents, enabling Claude to generate custom multi‑agent workflows—called harnesses—tailored to specific tasks.

Scope and Benefits

Dynamic Workflows are not limited to coding; they also support business‑plan analysis, resume screening, root‑cause investigation, and support‑queue triage. By using fan‑out, adversarial verification, and tournament mechanisms, Claude can split complex tasks among independent agents, reducing agent laziness, self‑preference bias, and goal drift.

How Dynamic Workflows Operate

A workflow runs a JavaScript file containing special functions that generate and coordinate sub‑agents. Standard JavaScript utilities such as JSON, Math, and Array are available for data handling. The workflow can select the model for each sub‑agent, decide whether the sub‑agent runs in its own worktree, and resume from the point of interruption if the session is paused.

Why Workflows Are Needed

When Claude executes a task in a single context window, prolonged processing leads to three failure modes:

Agent laziness : the agent stops early and claims completion after only partial progress.

Self‑preference bias : the agent favors its own generated results when asked to evaluate them.

Goal drift : repeated compression of the context causes loss of original constraints and boundary conditions.

Workflows mitigate these issues by giving each sub‑agent its own context window and isolated objectives.

Dynamic vs. Static Workflows

Static workflows created with claude -p required covering all edge cases manually. With Claude Opus 4.8 and Dynamic Workflows, Claude can automatically author a custom execution framework for a given use case, eliminating the need for exhaustive static definitions.

Useful Patterns

Classification‑then‑action : a classifier agent routes tasks to appropriate agents.

Fan‑out‑and‑aggregate : split a task into many small steps, run each in a separate agent, then combine results.

Adversarial verification : a second agent challenges the output of the first according to a scoring rubric.

Generate‑and‑filter : produce many ideas, then filter, deduplicate, and return the highest‑quality ones.

Tournament : multiple agents compete on the same task; pairwise judging selects the winner.

Loop‑until‑done : keep generating agents until a stop condition (e.g., no new findings) is met.

Representative Use Cases

Migration & refactor : using Bun, a workflow rewrote a Zig project into Rust by spawning agents for each step (call‑site, failing test, module) and running adversarial reviews.

Deep research skill : a workflow fans out web searches, extracts sources, performs adversarial verification, and assembles a cited report.

Deep verification : a workflow extracts factual claims from a report, spawns agents to fact‑check each claim, and validates the provenance of sources.

Sorting : for large lists, a tournament pipeline performs pairwise comparisons rather than absolute scoring, improving reliability.

Memory & rule compliance : a workflow creates a validator agent for each rule in CLAUDE.md, ensuring consistent enforcement.

Root‑cause investigation : multiple agents generate independent hypotheses from logs, files, and data, then validators challenge each hypothesis.

Large‑scale triage : agents classify backlog items, deduplicate, and either attempt fixes or hand off to humans, with isolation zones for untrusted content.

Exploration & taste : agents generate design or naming options, a reviewer applies a scoring rubric, and a tournament selects the top candidates.

Evaluations (Evals) : spawn agents to produce outputs, then comparison agents assess them against a rubric.

Model & agent routing : a classifier agent decides whether to use Sonnet or Opus based on task complexity.

When Not to Use Dynamic Workflows

Dynamic workflows are token‑intensive and still maturing. Reserve them for creative or high‑impact tasks that cannot be solved with a single Claude instance; routine coding tasks often do not require the extra computational overhead.

Prompting and Operational Tips

Write detailed prompts that incorporate the patterns above for best results.

Use the “quick workflow” trigger for small, focused tasks.

Combine /goal with /loop for repeatable workflows such as triage or research.

Set an explicit token budget, e.g., “use 10k tokens”.

Save workflows via the ~/.claude/workflows directory or distribute them as skills with a SKILL.MD manifest.

Dynamic Workflows expand Claude Code’s capabilities and open a new frontier for multi‑agent orchestration; experimentation will reveal further best practices.