Urgent.News

What's breaking now, across thousands of outlets.

AI

Claude Code Workflows: Critical Analysis

Claude Code's Dynamic Workflows: Theory vs. Practice Dynamic Workflows The Workflow tool creates a script - it's a dialect of JavaScript that is executed inside a special runtime - the script gets wrapped in an async function, so top-level async / await and return are legal, while certain functions like Date.now() are illegal because they break resumability. The script has a few primitives that…

Claude Code's Dynamic Workflows aim to improve multi-agent orchestration, but the implementation falls short in practice. The Workflow tool creates a JavaScript script executed within a special runtime, which has limited functions and strict rules to prevent interfering with the resumability of the process. Workflows are broken into phases like design, implementation, and verification, serving as a presentational element rather than a functional one.

The three main functional primitives are agent, pipeline, parallel, and workflow. The agent function prompts an AI agent to perform a task, while the pipeline declares a sequence of stages for agents to progress through independently. Parallel dispatches agents across a single stage, waiting for their completion before moving on. Workflow allows calling one workflow from inside another, up to one level of nesting.

Claude Code orients multi-agent orchestration around Claude, not the Claude Code application. Once a script is launched, the main session Claude spends minimal tokens on oversight, contrasting with Agent Teams where Claude must manage all agents. This shift is vital, akin to hooks being qualitatively different from reminder prompts - they operate outside the agentic loop, preventing costly token expenditure and preventing reliance on a model to make complex decisions during long-horizon tasks.

However, workflows prove costly. Claude Code plugins enable the shipping of saved workflows as first-class components, but most resources opt for second-rate ultracode emulators or workflow ports to Codex. Out of top 20 GitHub search results for "ultracode," none ship actual workflows for Claude Code, with most being reimplementations for other platforms. This suggests Claude Code users are scarce despite apparent interest in workflows.

Claude's handling of workflows is problematic. The Workflow definition suggests using the session model by default and only overriding it when confident a different tier is more suitable. However, this advice is rarely followed in practice. Additionally, enabling ultracode disables token efficiency constraints, often leading Claude to use workflows needlessly. Finally, the definition conflates cost and time, blurring the lines between these essential factors in workflow optimization.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

More from Tuesday 1 September →