1. What “Dynamic Workflows” Actually Are

On May 28, 2026, Anthropic introduced dynamic workflows in Claude Code — a capability that lets a single session fan a task out across tens to hundreds of parallel subagents, verify the results, and return one coordinated answer.

The headline isn’t “more agents.” It’s who writes the orchestration. Instead of a human hand-building a rigid pipeline, Claude plans the work dynamically from your prompt, breaks it into subtasks, distributes them, checks the output, and iterates until the answers converge. Work that you’d normally scope in quarters can finish in days.

The unlock is the underlying model. Opus 4.8 is finally intelligent enough to author a custom “harness” for your specific task rather than just run inside a generic one — and that single change is what reframes the entire discipline of working with AI.

2. Watch: The Concept in Motion

▲ A visual walkthrough of harness design and parallel agent orchestration.

If you absorb concepts better visually, start here. The sections below turn the ideas into something you can actually run.

3. Harness vs. Prompt: The Paradigm Shift

For most of the LLM era, the lever you pulled was the prompt. Clever phrasing, few-shot examples, chain-of-thought nudges, and 50-step prompt chains held together by hope. It worked — until the task got big enough that no single context window could hold the plan.

The default Claude Code harness is built for coding, and it turns out to be useful well beyond coding because many tasks resemble coding tasks. But to hit peak performance on specific classes of work — deep research, security review, large migrations — Anthropic historically had to hand-build a bespoke harness each time.

Dynamic workflows generalize that. Instead of anyone hand-building a harness per scenario, Claude writes one on the fly for the task you describe, then a runtime executes it. The harness becomes a first-class, reusable artifact: a script you can read, rerun, edit, and share.

4. How a Workflow Runs Under the Hood

When a workflow kicks off, the loop looks like this:

Because coordination happens outside the conversation, the plan stays on track no matter how large the task grows. Progress is saved as the run proceeds, so an interrupted job resumes where it left off instead of starting over.

# In Claude Code — two ways to trigger a workflow
# 1) Just ask, with auto mode on
> "Create a workflow to audit this repo for auth bypasses"

# 2) Flip the effort menu to the Claude Code-specific setting
/config → effort → ultracode   # sets effort to xhigh; Claude decides when to spin up a workflow

5. The Six Core Workflow Patterns

Dynamic workflows aren’t one trick — they’re a small family of orchestration shapes Claude composes for you. These are the patterns Anthropic’s own engineers lean on:

The common thread: each pattern pairs parallel breadth with independent verification. That’s where the quality comes from — not from any single agent being brilliant, but from many agents checking one another.

6. ultracode & “Ultra Code Mode”

The fastest on-ramp is the Claude Code-specific setting ultracode, reachable through the effort menu. Turning it on sets the effort level to xhigh and lets Claude decide automatically when a task warrants a full workflow versus a normal pass.

For the best experience, enable auto mode first. From there you either ask Claude to “create a workflow” explicitly, or you flip on ultracode and let it self-select.

7. Case Study: Rewriting Bun in 11 Days

The clearest proof of what this unlocks at scale is the recent rewrite of Bun. Jarred Sumner used dynamic workflows to port Bun from Zig to Rust with 99.8% of the existing test suite passing — roughly 750,000 lines of Rust, and just eleven days from first commit to merge.

A project that would traditionally consume weeks of senior engineering time was compressed into a window measured in days — with humans staying in the loop at the review gates that mattered.

8. Getting Started in 5 Steps

9. Merits, Costs & Trade-offs

The honest summary: dynamic workflows reward teams that have done the unglamorous groundwork — solid test suites, clean CI gates, sensible tool allowlists, and a branching strategy that lets you review a big change before it lands. The model can write the harness, but the result still depends on how trustworthy your gates are.

10. The Modern Agentic Stack

Dynamic workflows don’t live in isolation — they sit at the center of a 2026 toolchain. Here’s how the pieces fit, and a couple of tools worth pairing alongside them:

11. Pre-Flight Readiness Checklist

Before you point a workflow at anything important, make sure these are true:

• Your repo has a real test suite the fix loop can drive to green
• CI gates are clean and meaningful, not rubber stamps
• Tool allowlists are scoped so agents can’t reach beyond the task
• A branching strategy lets you review a big change before it merges
• You’ve run on one directory first to benchmark token cost
• Auto mode is on and you understand the confirmation prompt
• Admins have (or haven’t) enabled workflows in managed settings

12. What Engineers Are Saying

Quotes paraphrased and condensed from Anthropic’s launch announcement; see official docs for full context.

13. Frequently Asked Questions

The Bottom Line

The teams that internalize harness design before the rest of the market are the ones shipping reliable, auditable agents at production scale. The rest are still chaining prompts by hand. The shift from prompt engineering to harness design is the defining developer skill of 2026 — and with Opus 4.8 and dynamic workflows, the tooling is finally here to practice it. Start small, trust your gates, and let Claude build the scaffolding.