In This Article
Key Takeaways
- Dynamic Workflows let one 'lead' Claude write a JavaScript script that orchestrates tens to hundreds of subagents in parallel, then checks their work before you see it.
- The breakthrough is moving orchestration out of the chat window and into code — Claude becomes the planner, not the dispatcher.
- A developer used it to port the Bun project from Zig to Rust: ~750,000 lines, 99.8% of tests passing, in about eleven days.
- The skill it rewards isn't 'do the task' — it's describing a task clearly enough that it can be divided, delegated, and verified.
Anthropic introduced Dynamic Workflows for Claude Code: a way for one AI to break a large job into pieces, hand those pieces to many AI agents working in parallel, and then check their work before giving you a single coordinated answer. It launched as a research preview alongside Claude Opus 4.8 and is available through Claude Code and the Claude API, including on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
This is a genuine shift in how AI work gets done, and the plainest way to understand it is to stop picturing a single assistant and start picturing a team with a manager. Let me explain the mechanism, show you the proof, and draw out the one skill it makes valuable for everyone — even people who never write a line of code.
What Dynamic Workflows are
Until recently, using an AI meant one conversation with one assistant. Useful, but limited — like asking a single very fast person to do an entire project alone. Dynamic Workflows change the shape. When you describe a big task, Claude plans dynamically, breaks it into subtasks, and fans the work out across subagents running in parallel. Results are checked before they are folded in, and you come back to one coordinated answer while your session stays responsive the whole time.
Each subagent runs with its own context window and its own tool access. They work simultaneously rather than in a slow single file. One mind becomes a coordinated crew.
The key idea: orchestration becomes code
Here is the technical insight that makes it work, in plain language. A dynamic workflow is actually a small JavaScript program that Claude writes for your task. A runtime executes that program in the background, spawning and coordinating the subagents.
Why does that matter? Because it moves the orchestration logic out of Claude's limited memory and into code. The script holds the loop, the branching, and all the messy intermediate results. Claude's own context holds only the final answer. As Anthropic puts it, Claude becomes the planner, not the dispatcher. That separation is what lets a single session coordinate hundreds of agents without drowning in its own notes.
The orchestra, not the soloist
Picture a conductor. The conductor does not play every instrument. The conductor decides who plays what, when, and how the parts fit together. Dynamic Workflows make Claude the conductor: it figures out the plan, assigns sections to specialist agents, lets them work at the same time, and then judges whether the combined result is correct.
The leap is not a smarter soloist. It is a competent conductor who can keep a hundred players in time.
This is why the change feels bigger than a typical model update. It is not that the AI got more clever in a single conversation. It is that the AI learned to manage many conversations toward one goal — a fundamentally different kind of capability.
The 750,000-line proof
The headline demonstration came from developer Jarred Sumner, who used dynamic workflows to port Bun — a widely used software project — from the Zig programming language to Rust. The result was roughly 750,000 lines of Rust, with 99.8% of the existing test suite passing, completed in about eleven days from first commit to merge. Hundreds of agents worked in parallel, each writing one file as a behavior-identical port of its original, with two reviewers checking every file.
Let me put that in perspective. A port of that size is normally a job for a team over many months. The "test suite" is the automated proof that the software still behaves correctly, and 99.8% passing means almost nothing broke. Whatever you think of the hype around AI, that is a concrete, checkable, independently reported result — not a demo video.
Single assistant vs. dynamic workflow
| Aspect | Normal chat | Dynamic workflow |
|---|---|---|
| Workers | One assistant | Tens to hundreds of subagents |
| Coordination | In the chat window | In a JavaScript script (out of context) |
| Execution | One step at a time | Many pieces in parallel |
| Verification | You check | Agents refute & converge, then you |
| Claude's role | Dispatcher | Planner / conductor |
Why agents that argue produce better answers
One of the most interesting design choices is adversarial. In a well-built workflow, some agents address a problem from independent angles while other agents try to refute what the first ones found. The run keeps iterating until the answers converge. That is how a workflow reaches conclusions a single pass cannot — by building disagreement and verification directly into the process.
This mirrors how good human teams work. You do not want five people who all think alike; you want some who propose and some who poke holes, until what survives is trustworthy. Dynamic Workflows automate that healthy friction.
The skill that suddenly matters more
Notice what these tools reward. They do not reward “being able to do the task yourself” — the agents do that. They reward being able to describe a task clearly enough that it can be divided into pieces, delegated, and checked. That is a management skill, and it is becoming one of the most valuable abilities in technical work. Practice it: for any big task, write down the pieces, which can run in parallel, and how you will know each one is correct. That habit makes you better with or without AI.
Agents that improve between runs
Anthropic also added a capability it calls Dreaming to its managed agents: a scheduled process that reviews past sessions, surfaces patterns, and curates the agent's memory, so the agent gets better between runs. In plain terms, the agent reflects on what it did, keeps what worked, and carries the lesson forward.
That is closer to how a good employee grows than to how a static tool behaves. It is a small but telling sign of where things are heading: not just smarter single answers, but agents that learn from their own history.
What it means for you
You do not need to orchestrate a hundred agents to take the lesson. The valuable skill is learning to think like a conductor — to take something large, see its natural pieces, decide what can happen in parallel, and define how you will check each part.
So practice it deliberately. When you face a big task at work or in study, resist the urge to grind through it linearly. Instead, write the plan first: what are the pieces, which are independent, and what does "done and correct" look like for each. That discipline makes you more effective on your own — and far more powerful when you hand the pieces to AI. The future belongs to people who can decompose and verify, not just to people who can do.
Learn to orchestrate AI, not just chat with it
Our bootcamp teaches multi-agent thinking and the build-and-verify discipline behind tools like Dynamic Workflows. Hands-on, five U.S. cities, June through October 2026.
See Our BootcampSources: Anthropic, “Introducing dynamic workflows in Claude Code” (claude.com/blog) and the Claude Code documentation; InfoQ, “Claude Code Adds Dynamic Workflows for Parallel Agent Coordination” (June 2026); reporting on Jarred Sumner's Bun Zig-to-Rust port (~750,000 lines, 99.8% tests passing, ~11 days). Figures reflect the publicly reported demonstration.
What dynamic workflows can't do (yet)
Honesty requires naming the limits, not just the wins. Dynamic workflows are powerful, but they are not magic, and treating them as such is how people get burned. They excel at work that can be cleanly divided into independent pieces — porting files, reviewing many documents, running the same analysis across many inputs. They struggle with work that is deeply interdependent, where every piece needs to know about every other piece, because that is exactly the kind of work that resists being split up.
They also cost more. Spawning hundreds of subagents means hundreds of times the token usage of a single conversation. For a one-off task that a single model could handle in one pass, a dynamic workflow is overkill — like renting a construction crew to hang one picture. The skill is not just knowing how to use them, but knowing when they are worth it. The Bun port was worth it because 750,000 lines genuinely needed parallel effort. Your weekly report probably does not.
Where this is heading
Step back and dynamic workflows look like an early version of something bigger: software that manages software. For most of computing history, a human wrote the orchestration — the loops, the branches, the "do this, then that." Now the AI writes that orchestration itself, dynamically, for the task in front of it. That is a meaningful shift in who holds the plan.
I expect this pattern to spread well beyond coding. Any work that can be decomposed, delegated, and verified — research, data analysis, content production, operations — is a candidate for the same orchestrate-and-converge approach. The companies and individuals who learn to think in terms of "how would I break this into parallel, verifiable pieces?" will have a real edge, because that is the question these tools are built to answer. The conductor's skill, in other words, is about to become a lot more valuable across a lot more fields.
Common questions
Do I need to be a programmer to use Dynamic Workflows? To use them at full power, yes — they run inside Claude Code, a developer tool. But the underlying idea is for everyone: break a big job into pieces, do the independent pieces at once, and check each piece before combining. That mental model helps anyone organize complex work, code or not.
How is this different from just asking Claude a question? A normal question is one conversation with one assistant. A dynamic workflow is one Claude writing a script that spawns many sub-Claudes, each working on a piece in parallel, with results checked and combined automatically. It is the difference between asking one fast person and managing a coordinated team.
Is the 750,000-line code port real? Yes. Developer Jarred Sumner used dynamic workflows to port Bun from the Zig language to Rust — roughly 750,000 lines, with 99.8% of the existing test suite passing, in about eleven days from first commit to merge, with hundreds of agents working in parallel and reviewers on each file.
What does “the context stays clean” mean? Normally an AI's limited memory fills up with all the intermediate steps of a big job. Dynamic Workflows keep the loop, the branches, and the intermediate results inside the script instead, so Claude's own memory holds only the final answer. That is how it can coordinate hundreds of agents without losing track.