The Claude Update Everyone Missed (Dynamic Workflows)

本站添加: 2026-05-31

[00:00:00]All right, so take a look at this. This is an entire due diligence report completed on 70 plus documents. And these documents include things like contracts, leases, proposals, memos. And I was able to spin up a team of 50 plus agents to go through all this documentation, create all the analysis, and then synthesize this in this easy-to-read report. And I was able to generate all this in the span of 20 to 30 minutes, where the average person would take hours. And what it did behind the scenes is go through these 70 plus different files and folders, and then spin up all the agents not to do all the work in tandem and all execute things in parallel, but rather work together. And each one would come with its own set of insights that would then be cross-checked and validated by other agents. So, it's not about creating agents for the sake of agents. It is combining all the resources and combining this brand new feature, which most people are still not talking about as much as they should. Even though Opus 4.8 has taken the headlines, it's really a small step change from the prior version. The real feature you should be paying attention to is the dynamic workflows feature that I'm going to walk you through in this video. This is meant to be your secret weapon for incredibly large-scale tasks that require more than five or six agents at a time. So, in this video, I'm going to walk you through exactly what dynamic workflows are, how they work, and how you can apply it to actual practical use cases.

[00:01:26]Let's jump in. And before we dive deeper, I want to show you how you can quickly use the workflows function. You have a couple different ways you can do it. The first way is doing {slash} workflows. And you'll only see this if you update to the absolute latest version of Claude Code, and then you'll be able to add an argument here. This could be whatever you're trying to execute in the full scope. Now, the next way is it's very similar to how you can invoke ultra think, then you get these pretty colors that basically run very deep reasoning for one specific prompt.

[00:01:54]You can now do the same thing with workflows. So, if you say something like I want to create a series of workflows that goes through my entire data room and then inspects all documents that don't have proper disclosure, then you should be able to see, there we go, that we have these extra colors right here denoting that it will be using the workflows function to create these teams of agents to execute this core task. Now, one additional thing I noticed is that if the task lends itself to using workflows, once in a while Opus will actually default towards using this function as well. Now, to give you a preview on how you can write a prompt to invoke this for actual practical use cases, this is the very prompt that I used to generate that entire due diligence report that you saw in the intro. So, it starts off saying, "This is a folder data room for a company we're acquiring called Northwind Logistics." Again, this is all hypothetical. "I'm the diligence lead.

[00:02:46]Build a workflow." This is the key right here. This is what tells Cloud Code to use this. "That reads every contract here in parallel and flags anything that could hurt the deal. Change of control clauses, etc., etc." A lot of legalese language. So, this gives the entire scope of what we're trying to accomplish. And then it will go through all the documentation, summarize, and then decide how many agents it needs to execute this. And you can see at the very top of this section here, you'll see it starts to use this function workflow. And then this is where it defines the scope of the workflow, what it has to use, how many agents it might need, what those agents would be doing, and then it creates the entire specification of how this will run. Now, I decided to dive a little bit deeper and have these teams of agents build another workflow to do an incredibly detailed deep dive to look for little needles in the haystack in all of the data folders. So, I said the following, "Build a workflow that reads every file in this entire data room and hunts things that a seller would rather we never find." So, again, looking for micro fine print. So, this will go through all of the things that I'm looking for. And this one, because of the level of detail and the fact that we have to go through every single one of these folders, each of which has a series of very detailed documents, this takes way longer. This spun up 51 agents, and it ended up taking almost 3 million tokens. Now, did we end up with a good result? Did we end up with the walk-through that we were expecting?

[00:04:18]Yes. I'm obviously not going to review it on this video, but walking through all the files, it did take a fine-tooth comb across each and every file. But, if we take a look at this screenshot here, this did end up taking 23 minutes and burning close to 3.2 million tokens for this process. Now, depending on how well positioned your prompt was and what the output looks like, it could be worth it.

[00:04:40]But, you want to be very careful because this is not something that is meant to be used every single day. And by the way, if you want to go infinitely deeper on things like Claude Code and Codex and how to take your agentic systems to the next level, then you want to check out the first link down below for my early adopter's community. We drop a community-driven module every single week to keep answering the most pertinent questions we get every day.

[00:05:01]So, if you're interested, check that out. All right, back to the video. So, again, breaking things down, you always want to use the keyword workflow here.

[00:05:08]And it really starts with that core instruction. From that instruction, it will start to spin up a series of functions, and then by proxy, a series of agents. And the agents, like I said, aren't meant to just work on their own.

[00:05:20]They're meant to create some form of outcome or derive an insight depending on the task at hand, and then have other agents kind of spar with them to decide, is this a legitimate insight? Does this deserve to end up in our final output?

[00:05:33]And the natural inclination from anyone looking at this feature is, this is yet another way that Anthropic is trying to push you to burn as many tokens as possible so you pay more. And my pushback to that would be, this is not designed for a very particular task that is generic. This is for something that you know has to go through an audit every single part of a repo, a code base, and look through and be very comprehensive in its analysis.

[00:06:01]And theoretically, if you were to spin up a series of agent teams over and over again to accomplish the same task, one agent team costs around 250,000 to 300,000 tokens on average for one particular task. Because those agents are working together, they're spinning up the equivalent of a Kanban board, looking at tasks in to do, in progress, and completion. And you could do this three or four times, which would bring you to above one to 1.5 million tokens.

[00:06:29]So, this is a more efficient way to leverage agent team dynamics and agentic patterns within a very comprehensive and compressed way. And the way it works is they don't have 51 agents spin up at the exact same time. All of the agent specifications, roles, and responsibilities are specified, and then they are queued up, and then as core agents complete their tasks, then in batches, the next ones are spawned to either check the work of the prerequisite agents or to move forward with the next tasks. So, running something like this becomes more worth it when you have a very repeatable, large-scale workflow. It's very wide in its scope, and it's worth actually pushing forward till the very end. So, you don't want to stop at, let's say, 70 folders. You don't want to have Cloud stop at 40 and then spin up a brand new agent team to complete the task. You want this to be a long-running task with a very defined beginning, middle, and end where it can run pretty much autonomously once it has all the prerequisites it needs. And if you need a mental picture of how this looks like on the inside, imagine you have agent A, and agent A looks through, let's say, data and comes up with a result or an insight. Agent B will have to push back on it and double-check and play devil's advocate. So, you basically have these micro devil's advocates running in the background to make sure that by the time you get a confirmed insight, it is indeed as meritorious as possible and ideally as data-backed as possible. In terms of hands-on use cases, you could think of one for each and every industry. And I want to go through and speedrun a few of them. So, imagine in law, like you saw in the data room example, you could read every contract in a data room and flag all the deal killers before you actually enter a meeting. So, your prompt would look something like, "Build a workflow." I'm going to keep harping on this to make sure you know that this is the trigger that reads every contract in this folder, flags change of control, auto-renewal, uncapped indemnity, a series of other legal-ease things here, and then you would basically enter that.

[00:08:26]And ideally, if you want to provide more context about the case, you would throw that in. Or, ideally, you might even go into plan mode before you actually execute the workflow to make sure that you have the entire plan filled out on exactly what the critical path will be for all these agents.

[00:08:42]When it comes to finance or the finance function, you could do something like scanning thousands of transactions from things like QuickBooks, CSVs, or wherever your financial data lives. And your prompt would look something like, "Build a workflow that reviews every transaction in this export, flag duplicates, outliers, miscategorized entries, verifies each flag against the source rows." Source rows would be basically telling me which rows of data approximate to this data anomaly that you found, and then return a clean exceptions report. Next up, we could have healthcare. So, you could then check every patient chart against the latest care guidelines in one pass.

[00:09:19]Naturally, you want to make sure that you're running this in a secure manner.

[00:09:21]I wouldn't imagine you're using a consumer-based cloud code account. You'd need the Amazon Bedrock passthrough for enterprise usage. And then, by then, you'd be able to enter a prompt that says, "Build a workflow that checks every chart in this folder against the guidelines in guidelines.md, flags gaps in documentation or follow-up, verifies each gap, and returns a prioritized review list. For insurance, having worked in it myself as an intern before, you could do something like triage a mountain of claims and route only the ones that need a human. So, theoretically, you could sift through and tag the ones that are more formulaic versus the ones that actually require further human review. Your prompt could be build a workflow that triages every claim in this folder, sorts them into two folders, auto approve and needs a human, and then explain why it thinks it needs a human decision from the ones it tagged, and then it would return a queued list by urgency, so a priority list. In real estate, you could use this to abstract an entire lease portfolio into one clean table overnight. So, if you have a bunch of lease contracts and you want to be able to go through them and consolidate them, you'd be able to run a workflow that says, "Abstract every lease in this portfolio, pull rent, renewal, term, and key dates into one table, verify each extraction against the document, and flag anything unusual." So, especially when it comes to anomaly analysis, looking for weird things amongst large swaths of data, this becomes a very helpful tool, and in many cases, it's worth the token burn.

[00:10:52]And actually, given that it's Claude Code, this feature is primarily designed for software engineering, even though I'm showing you more business-oriented use cases. In software, naturally, you could ask it to go through a very large codebase to find one bug hiding across hundreds of files. So, I just actually used this morning for one of my more enterprise codebases using all kinds of AWS, Amazon Web Services, microservices, and it probably has 7 to 10,000 lines of code. So, every single time I load it up, I'm very specific with what files I give Claude Code access to to increase the likelihood that I can use my entire context window. But, if I want to run a security review, if I want to look for any key gaps or open ports in this application that's deployed online, this would be an interesting way to run all of these agents to battle test and look for every single gap that my existing security reviewer might not catch. So I prompt could say build a workflow that audits every file under SRC for the missing authorization checks, has a second agent try to refute each finding, and return only the confirmed issues with the file and the exact line. So in case you don't want to make any changes to the code, but all you want is a review of where to review said code for a human in the loop developer to take a look and see is it worth actually changing this, then this could put it on a silver platter for you. Now for marketing folks, you could use this for competitor analysis. So let's say you lined up a list of a hundred of your closest competitors in your country, your region, etc. You could write a prompt that says audit each competitor site in this list for messaging, offers, and SEO gaps, and then cross-check the findings and return one overall competitor sheet. Now, could you run this without this feature? Potentially.

[00:12:37]What if the number becomes a thousand or three thousand? An agent team or a series of sub-agents would struggle to accomplish this task. Now for our second last example, when it comes to recruiters, you're dealing with hundreds if not thousands of applications that you have to sift through or you rely on your existing application tracking system to go through. Instead of using that ATS that could be flawed, if you have a certain criteria that you want to be able to filter through and not have it offloaded onto another system, then you could say build a workflow that scores every resume in this folder against the rubric in, let's say, a markdown file that you put together potentially using AI, and has another agent, or in this case many agents, check each score for bias inconsistency and return a ranked shortlist with reasons and rationale. So now instead of just getting this spitted out list from your ATS system, which is basically looking for keywords or semantic search, you now have something that's a lot richer. And last but not least, when it comes to the thorny topic of compliance, you could write a prompt that says, "Build a workflow that checks every policy in this folder against the standard in, let's say, some markdown file you put together, flags each gap, has a second agent or series of agents, refute false positives, and return a ranked remediation list." So, hopefully you can get the gist now. Big tasks, many documents, large scope, ideally all done in one shot. And that's pretty much it. Hopefully this gives you a good walk-through of what dynamic workflows are and when it might make the most sense for you to use them. If you want access to all the prompts I showed you right now, on top of the ones I showed you at the very beginning, you can grab all of them completely for free in the second link down below. And for the rest of you, if you found this to be a hype-free walk-through of this feature, and you want to see more content just like this, then please leave me a like on the video, and if you so choose, a comment would be super helpful as well.

[00:14:29]I'll see you in the next one.