Big Idea

The argument in one line.

Dynamic workflows fix the three ways Claude degrades on complex single-session tasks by routing work through isolated sub-agents with fresh context windows, and UltraCode is simply the automatic mode that decides when to invoke them.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code regularly and have hit the ceiling where it stops partway through a large multi-part task.
You run deep research, large migrations, or codebase-wide bug hunts and want adversarial verification built in.
You want to understand the practical difference between UltraCode, /effort, /workflows, and agent teams.
You are on a Claude Max plan and want to know what a real dynamic workflow run actually costs in tokens.

SKIP IF…

You are not on Claude Code -- this feature is Code-specific and not available in the Claude web app.
Your day-to-day tasks are routine; the token overhead is not justified for straightforward prompts.

TL;DR

The full version, fast.

Claude Code defaults to a single context window for every task, which degrades on complex work through three documented failure modes: agentic laziness, self-preferential bias, and goal drift. Dynamic workflows combat this by spinning up isolated sub-agents with their own context windows. UltraCode is the /effort level that bumps effort to xhigh and enables automatic dynamic workflow orchestration. A live deep research demo ran 101 agents, consumed 3.7M tokens in 11 minutes, and cost 4% of a 200-dollar-per-month max plan.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:40

01 · Hook

UltraCode claim stated upfront: hundreds of agents, custom harness built on the fly. Promise to break down exactly how it works.

00:40 – 05:42

02 · UltraCode Explained

Static vs dynamic harness comparison using a checkout migration example. What /effort ultracode does: effort to xhigh plus automatic dynamic workflow orchestration. Difference between forcing /workflows manually vs UltraCode automatic mode. Self-sponsored masterclass plug at 3:10.

05:42 – 09:51

03 · Dynamic Workflows Deep Dive

Anthropic blog walkthrough: three failure modes of single-session complex tasks (agentic laziness, self-preferential bias, goal drift). Six workflow patterns visualized. Docs comparison table: subagents vs skills vs agent teams vs workflows.

09:51 – 14:37

04 · Demo

Two live demos. Demo 1: /deep-research on dynamic workflows itself -- 5 phases, 101 agents, 3.7M tokens, 11 min, 4% of 200-dollar max plan. Demo 2: parallel bug hunt on Next.js app -- 34 confirmed bugs, 7 false positives, adversarial verifier per finding.

14:37 – 15:13

05 · Final Thoughts

Summary of key points. Token cost acknowledgment. Chase AI Plus CTA.

Atomic Insights

Lines worth screenshotting.

UltraCode does exactly two things: bumps effort to xhigh and enables automatic dynamic workflow orchestration -- it does not change the model you are on.
The static harness failure is architectural, not an intelligence problem: one context window accumulates compaction loss, self-grading bias, and partial-completion acceptance.
Self-preferential bias means Claude in the same session cannot reliably verify its own work -- adversarial verification requires a separate agent with a fresh context.
Dynamic workflows are available without UltraCode: /workflows forces a workflow for any prompt; UltraCode makes the decision automatic based on task complexity.
Saved workflows are reusable like skills and can be repeated across sessions, not just invoked once.
A 101-agent deep research run cost 4% of a 200-dollar-per-month Claude Max plan -- about 3.7M tokens in 11 minutes.
The bug hunt demo ran in half the time and used half the tokens of the deep research demo, suggesting workflow cost scales with information-gathering breadth more than task complexity.
Anthropic ported the Bun runtime from Zig to Rust in about a week using dynamic workflows, validating the feature on a real large-scale migration.
The adversarial verifier agent catches false positives and provides per-finding evidence, which a single-session review cannot do reliably.
The /deep-research slash command is a preloaded dynamic workflow that ships with Claude Code -- available immediately on an updated install without custom setup.

Takeaway

Three failure modes that break every complex Claude session.

WHAT TO LEARN

Claude does not degrade on hard tasks because it lacks capability -- it degrades because single-context-window architecture creates three compounding problems that sub-agents eliminate.

Agentic laziness is structural: when a task is large enough that the model loses track of completion criteria, it will declare done at partial progress rather than continue.
Self-preferential bias means you cannot ask Claude to verify its own work in the same session -- the verification will favor the original output regardless of actual quality.
Goal drift accumulates silently across many turns; each context compaction step is lossy, so constraints present at turn one may be missing by turn forty.
Dynamic workflows fix all three by giving each sub-agent an isolated context window with a narrow focused goal -- no accumulated bias, no drift, no partial-completion shortcuts.
The token cost of a dynamic workflow is front-loaded, not additional: 101 agents consumed 4% of a 200-dollar-per-month max plan for research that would have required multiple manual sessions otherwise.
The adversarial verifier pattern is the most practically valuable element for any task where false confidence is the main failure risk.
Saved workflows are reusable: a workflow built for one task can be stored and run again, amortizing the orchestration cost over repeated use.
The preloaded /deep-research workflow ships with Claude Code and requires no setup -- it is a working example of fanout-and-synthesize available immediately on an updated install.

Glossary

Terms worth knowing.

Static harness: The default Claude Code approach: a single context window, sequential tool calls, no sub-agents. Works for most tasks but degrades on large complex ones.
Dynamic workflow: A custom orchestration script Claude generates at runtime for a specific task, spinning up multiple sub-agents with isolated context windows rather than working in one session.
UltraCode: A Claude Code effort level set via /effort ultracode that bumps effort to xhigh and enables automatic dynamic workflow orchestration.
Agentic laziness: A documented failure mode where Claude stops before completing a complex multi-part task and declares the job done after only partial progress.
Self-preferential bias: The tendency to favor earlier results when asked to verify or judge them against a rubric within the same context window.
Goal drift: Gradual loss of fidelity to the original objective across many turns, especially after context compaction, where summarization steps drop details and constraints.
Adversarial verification: A workflow pattern where a separate verifier agent independently checks the findings of worker agents, catching false positives that self-review would miss.
Fanout-and-Synthesize: A workflow pattern that spawns many parallel agents to gather from multiple sources, then a synthesis agent cross-references and produces a final report.
Classify-and-Act: A workflow pattern where a classifier agent routes each subtask to the most appropriate specialized sub-agent.

Resources

Things they pointed at.

05:43linkAnthropic blog: A harness for every task -- dynamic workflows in Claude Code ↗

09:01linkClaude Code docs: Dynamic workflows ↗

03:13productChase AI Plus Claude Code masterclass ↗

Quotables

Lines you could clip.

06:07

“The longer Claude works on a complex task in a single context window, the worse it becomes.”

Tight declarative claim, no setup needed, backed by Anthropic documentation→ TikTok hook↗ Tweet quote

12:12

“101 agents, 3.7 million tokens in eleven minutes later, the workflow is complete. In terms of usage on my weekly max plan, it was 4% and I am on the 200-dollar-a-month max plan.”

Concrete numbers address the token cost concern immediately→ IG reel cold open↗ Tweet quote

14:52

“Before this, it was kind of hard. We are bringing in outside orchestration layers, and now it is all ready to go inside of Cloud Code itself.”

Clean before/after framing→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogy

00:00This is the most powerful Claude code update in months, and nobody's talking about it. UltraCode allows Claude code to handle extremely large complex tasks by spinning up an army of agents. And I'm not just talking about a handful.

00:13I'm talking about potentially hundreds of agents to split up the work and get it all done in one go. But the real power of UltraCode isn't the amount of agents it can spawn. It's the fact that it builds a custom harness tuned to your specific task on the fly.

00:27Normally, Claude code takes a one size fits all approach to solve your problem, but with this brand new feature, that's no longer the case. So in this video, I'm gonna break down exactly how it works so you can start leveraging this new feature right away.

00:40So why should you care about UltraCode and dynamic workflows? Well, it's pretty simple. It is the best way to handle large complex tasks with Cloud Code.

00:49That's it. And the way it does it is through custom harnesses.

00:54That's a fancy way of saying it's gonna create a novel way to solve whatever it is you're trying to solve. A great example is how it handles something like deep research. If you have a rather complicated question, like, should we migrate our checkout service to a new provider?

01:10If I'm just using Claude code as it normally operates, I just hop inside the prompt window and I give it this prompt. Well, it's gonna use its static default harness.

01:21It's gonna be in one single session. Right? We're gonna use one context window.

01:25It's gonna do a handful of web searches. It's gonna fetch the top results. It's gonna verify, summarize, and give us a generic research report.

01:33By default, when we talk about static harness, what we're really saying is you just open up Cloud Code here in one single context window, and we're not telling it to use sub agents or do any sort of, like, antagonistic review of its original thought process. So it's rather simple.

01:47It uses less tokens, and in the end, you get a generic answer. Caveat, that's fine for most issues. We're talking here today about large complex stuff.

01:55By contrast, if we use Ultracode and then dynamic workflow, and I'll talk about in a second the difference between those two, we get a custom harness that is built for that problem. We are gonna have a different way of solving that issue. We're not just gonna sit in one context window.

02:10We're gonna bring in different agents and the different agents are gonna do different things, again, depending on the task. So in this example, where I say, hey. Should we migrate our checkout service to a new provider instead of just doing some web searches and coming up with a summary?

02:24Instead, with a dynamic workflow, it's gonna be like, okay. Well, what is our checkout service? So it's gonna read our billing code.

02:30It's actually gonna go deep into how we actually operate. From there, it's then gonna check the features against the new provider's documentation. It's gonna price our transaction volume.

02:40It's gonna have an actual devil's advocate agent to say, does that actually make sense? And then in the end, we get a specific recommendation instead of a generic one.

02:50So the big thing with dynamic workflows is, again, the process by which we come up with an answer is different. It's customized, and it depends on the question we are asking.

03:01And by doing it this way, we get better results. That's why you should care. Now before we talk about where UltraCode plays into all this, first, a message from today's sponsor, me.

03:12So not too long ago, I just released a Claude code masterclass, and it is the number one way to go from zero to AI dev, especially if you don't come from a technical background. I update this every week. I've recently added modules for a codex masterclass as well.

03:25And if you wanna get your hands on this, you can find it inside of Chase AI plus. I will put a link to that down in the pinned comment. So what is UltraCode?

03:32They must have cared about it to give it this fancy graphic when you bring up forward slash effort. So UltraCode is related to effort level.

03:40So if you do forward slash effort inside of Claude code, you will see the spectrum that pops up from low all the way to UltraCode. Normally, by default on Opus 4.8, we're on high, but UltraCode goes kind of a step beyond max.

03:53Well well, kind of. What actually happens when I do forward slash ultra code, two things occur.

03:59One, my effort level goes from high to extra high. We're not on max. We're just jumping to extra high.

04:05And secondly, I now have automatic dynamic workflow orchestration.

04:10So there's two things we're talking about here. Right? In this video, we've talked about dynamic workflows, which is, you know, this idea that we're doing custom harnesses for our task, and we have UltraCode.

04:20Okay. So UltraCode changes effort to extra high and allows for dynamic workflow orchestration automatically.

04:28Dynamic workflows are also its own thing inside of Claude code. So if I'm inside a Claude code and I do something like forward slash workflows, I'm now forcing Claude code to create a workflow for whatever prompt I give it, you know, like insert task.

04:44But if we're in UltraCode, which we are now, Claude code will on its own decide, hey. Does this need a dynamic workflow or does it not? So we have the option to always invoke dynamic workflows if we just do forward slash workflow or if I say something like, hey.

04:59Use workflows. So it's it's it's almost like a skill that I can invoke, same sort of deal. Or UltraCode allows Cloud Code to be kinda smart about it, and it's gonna be like, hey.

05:09Depending on the prompt, sometimes we'll just do the static harness. We don't always need to go crazy. Or if it's complex enough, hey.

05:15I'll go dynamic workflows. So that's where UltraCode comes in. It's like it's like an automatic layer.

05:20You don't have to think about it. If it needs dynamic workflows, it will do it. You don't have to use your brain.

05:24Right? And that's great. So with that being said, to actually maximize UltraCode, we need to dive a little bit more into dynamic workflows because you understand the why, you understand UltraCode versus dynamic workflows.

05:37Now now let's talk about this a little bit more. And CloudCode actually Anthropic wrote an entire blog about this. So the blog which came out about a week ago is called a harness for every task, dynamic workflows in Cloud Code.

05:48We're not gonna dive into this whole blog in this video. I'm just gonna hit the parts that you need to know about, and I'll put a link to it down in the description as well. Now the first thing I wanna highlight is why dynamic workflows?

05:57We touched on it at the beginning. Hey. It's better for complex task, but why is Claude normally not good enough for this thing?

06:02Well, that is because the longer Claude works on a complex task in a single context window, the worse it becomes. And the three things they talk about here can kinda all fall under the context rod umbrella, the idea of agentic laziness.

06:16You know, we've all run into this where you tell Claude code to do something that's rather large in scope, it kinda does some of it. We also have self pre preferential bias referring to Claude's tendency to prefer its own results or findings, especially when asked to verify or judge them against a rubric.

06:31Hey. If you remember my video a couple days ago where I showed you grill me codex bringing codex into the equation, this is something I talked about there as well. Right?

06:39Claude isn't great at evaluating its own work, if you're telling it to evaluate work within the same session. Like, you're in the same context when you're like, it's not graded it.

06:49And then lastly, goal drift. And again, keep beating this drum. Complex task, goal drift is gonna be a big thing.

06:55We have to handle this. We can't do these things in one session. And so creating a workflow, a dynamic workflow helps combat these by orchestrating separate Claude sub agents with their own context windows and focused isolated goals.

07:08Right? Sound familiar? GSD, superpowers, all these things these last few months are all coming to this one point of, like, how do we handle big tasks with the unlimited context window?

07:19It all ends up coming back to, like, sub agents, fresh context windows, that sort of thing. Now the last thing I'll touch on from this blog is some of the workflow patterns they bring up. This is not exhaustive.

07:29There's an infinite amount of workflow patterns, but I think it helps when we talk about custom harnesses and talk about dynamic workflows to see what we're actually speaking of, like, visually. Now the first example they give is classify in act. That's a workflow pattern where we have some sort of task.

07:43That task involves a number of subtasks, and we wanna divvy those out to appropriate sub agents. To do that, we need a classifier and using a dynamic workflow.

07:52Cloud Code will already know this is the big part. Cloud Code will already know that this is the sort of workflow pattern you need. Right?

07:59It will automatically set this up for you. Another one is fan out and synthesize. Think of deep research.

08:03Hey. I want you to research some sort of thing. I need you to go out there and get a ton of information from a ton of different sources, potentially hundreds of sources.

08:11I need you to bring them in. I don't just want you to summarize. I want you to verify.

08:14I want you to actually, like, cross reference it. And then at the end, I want a final report. Again, if you ask Claude code using UltraCode or using dynamic workflows to do deep research on some sort of thing, well, this is the sort of workflow we come up with.

08:27And it continues down the line. Adversarial verification, loop until done, a tournament style thing where you have a bunch of different ideas and judges, and at the end, you get a final winner, generate and filter.

08:37The blog itself goes into detail on all these, but at least for me, it was helpful to kinda see what they're talking about. I'm like, alright. Custom harness.

08:44What does that exactly mean? Well, it just means there's a number of paths to find a solution.

08:50And with a custom dynamic workflow, we get all these options. And the point is they're gonna be custom fit to the task versus, you know, static harness, do a web search, ask a couple questions, summarize.

09:03We can do better. And lastly, before we go into the demo, I'm gonna highly suggest that you take five minutes of your time, go to the Cloud Go docs, and actually read what they have written up about dynamic workflows so you can get a better sense of how it's working under the hood versus something like agent teams. Hint, it's actually running a script at runtime execution and things like how to save workflows because you can repeat them.

09:24They're kind of like skills in that sense and that sort of thing. They they actually have a really good write up here. And in this write up, they actually tell us about a preloaded dynamic workflow that comes with Claude code.

09:35It's actually a deep research workflow similar to the deep research that's been around for a while on the actual web app. So we'll do two demos.

09:42The first one I'm gonna show you so you can kinda see how this all works is this deep research one. And this is inside your Claude code right now if you're updated. All you have to do is do forward slash deep research.

09:51So inside a Claude code, we're gonna do forward slash deep research, and then I'm just going to give it a prompt. Well, actually, it'd be kind of a meta prompt. We're gonna have it do deep research on dynamic workflows and the best practices for creating them.

10:05Can you do some deep research on the brand new dynamic workflows and UltraCode within ClaudeCode? I wanna report detailing the best practices. There's a lot of talk about there being custom harnesses with the dynamic workflows.

10:20How do we make sure the custom harness that's built using dynamic workflows is the best one for the job, or is that something on the user level we just expect Cloud Code to handle? So there we go.

10:31So when we did that, you see we got a couple messages. Running deep research workflow, topic clear enough, proceed, workflow, dynamic workflow, deep research, and then it's telling us the workflow has been launched in the background.

10:43It has five phases, scope, search, fetch, verify, and synthesize. And we have the ability to watch it live. So if I go ahead and do forward slash workflows, you can see right here all of the agents is what is happening in real time.

10:58And so for the scope, we just have a single agent. For search, we will have five agents. And as these start working, we'll we we can actually oops.

11:06Once these start working, we can actually see their token usage because one of the big costs of this right. You can kinda see it right here. One of the big costs we need to think about when it comes to using dynamic workflows in UltraCode is the token cost.

11:19It is token heavy. Now there's definitely an argument to be made that we're just kind of front loading the cost and that by virtue of us having more effective results from UltraCode and dynamic workflows, we're probably saving tokens in the long run.

11:34But just don't be surprised, especially if you're using a dynamic workflow that I wasn't joking at the beginning is using a 100 plus agents. Don't be surprised if you have a wild token cost at the end of that.

11:45So we can see here for our five agents that are searching right now, they've used about 250,000 tokens a pop. Scope itself took about 40 k.

11:54And then fetch over here, it looks like it's has potentially 12 sub agents. And then we also have a list up here. Right?

12:01Four out of 22 agents and two minutes have. And I'll also reference my total weekly usage at the end of this as well and how much I burned. So 101 agents, 3,700,000 tokens in eleven minutes later, the workflow is complete.

12:16And in terms of usage on, like, my weekly max plan, it was 4% and I'm on the $200 a month max plan. So, you know, token usage stuff, it is no joke.

12:24Like, have to know when to apply this. And then I had to go ahead and turn the report into this HTML asset you see here, and it's pretty much restating a lot of what we see in the actual Cloud Code documentation involving dynamic workflows.

12:39So what are some other use cases for dynamic workflows in UltraCode? Well, Anthropic actually spells it out for us. Things like code based wide bug hunts, large migrations, and critical work that needs to be checked twice.

12:50Another example Anthropic brings up is rewriting Bun with dynamic workflows where they actually ported Bun from Zig to Rust over the course of about a week using this feature.

13:03Let's test it out with a bug hunt. I'm inside the directory for my AI agency website, which also doubles on the back end on an admin side.

13:10It's sort of like my content creation command center. So we're going to have it run a dynamic workflow to sort of do a bug hunt. Can you go ahead and use dynamic workflows inside this directory to run a bug hunt and see what sort of bugs we have in this current directory.

13:30Once you create the report, turn it into an HTML report and bring that up in my browser. So you can see right here calling the workflow command cluster parallel bug hunt across the Next. Js app.

13:40Adversarially verify each finding, synthesize a severity ranked report. So it's running in the background, and as always, we can do forward slash workflows to take a look.

13:50So for this bug report, it ran in about half the time and took half the amount of tokens as deep research. We had 34 confirmed bugs. It had seven bugs that were false positives.

14:00And of those 34, two are high, nine are medium, and 23 are low. I have the ability to click on them. It shows me what's wrong, the evidence, the fixed, and then also the adversarial verifier that's saying like, This actually is a bug.

14:14Here's the problem. And it gives me this report for pretty much all of them. Where the issue is, what it kinda defines the error as, what's wrong, evidence fixed, adversarial verifier.

14:24So it's pretty deep in terms of what it's actually finding. And I think the best part of this is the adversarial verifier. Because again, one of the big issues with Claude coding complex task is like, can you confirm that it actually did what it's supposed to?

14:37So that's where I'm gonna leave you for this video. We went over why you should care about UltraCode in dynamic workflows, how they work, and went through a few examples. I think this is an awesome feature.

14:46I think it's extremely powerful. Yes.

14:49It's extremely token heavy, but sometimes we do need the big guns, especially for tasks we really, really care about. And before this, it was kind of hard. Right?

14:58We're doing some hacky things. We're bringing in outside orchestration layers, and now it's all ready to go inside of Cloud Code itself. So as always, let me know what you thought.

15:07Make sure to check out Chase AI plus if you wanna get your hands on my Cloud Code masterclass, and I'll see you around.

The Hook

The bait, then the rug-pull.

A new effort level inside Claude Code quietly shipped that turns a single-session prompt into a custom-built orchestration of hundreds of isolated agents. The creator calls it the most powerful update in months -- and the live demo numbers back it up.

Frameworks