AI Agents for Lawyers: The Sub-Agent System That Multiplies Your Output

The productivity gap that I'm noticing inside of law firms and legal teams that are all using AI is less about the platform or the AI model that they decided to adopt, and it's more about

the capabilities of their team members, the paralegals, their lawyers, their associates. And specifically, a real

associates. And specifically, a real leverage point is whether the users of AI know how to run more than one thing at a time. What if you could run contract review, conflict check, and a multi-jurisdiction research

simultaneously?

Not one after the other, but all at the same time. Just imagine the productivity

same time. Just imagine the productivity gains that you can get from doing all those things at the same time. Anthropic

actually ran a head-to-head test that had one Claude agent against a team of Claude agents, and the team won by 90%, meaning they completed the task 90% quicker. And that gap exists in your

quicker. And that gap exists in your legal workflows right now. There's an

opportunity there to get more done with the same amount of effort. So, we

custom-build AI automations for law firms and in-house legal teams. We also run workshops where we coach lawyers, attorneys, paralegals how to use Claude, Gemini ChatGPT in their day-to-day. And one of the

skills that we've been teaching more and more of late is the concept of sub-agents and agent teams. And so, in this video, I'm going to show you what sub-agents actually are, why they outperform a single agent on specific

legal tasks, and just as importantly, which legal tasks you should never run sub-agents and agent teams on. I also

built a free tool to go with this video.

It takes any legal task you're thinking of automating and tells you exactly how to structure it as a sub-agent workflow.

Watch till the end of the video and I'll tell you exactly where you can get it.

And by the end of this video, you'll know how to use sub-agents, agent teams, when not to, and how to map your own workflows into this system. And you'll

have a tool that can actually do it immediately for you.

Now, here's the thing. When people use Claude Co-work, they're doing this.

They've got one agent going, they've got one conversation, and one task at a time. And for a while, it really does

time. And for a while, it really does feel and it is a real productivity gain.

But as you become more accustomed with using AI in your day-to-day, and you've grasped some of the fundamentals and basics such as good structured prompting, you're going to want to do more. You're going to want to get more

more. You're going to want to get more done in the same period of time. And

that's where sub-agents come in. They

essentially flip that one agent one conversation model, and they flip it on its head. So, instead of one agent doing

its head. So, instead of one agent doing everything in a sequence, you have what we call an orchestrator or a manager agent. Think of it like a senior

agent. Think of it like a senior associate who is managing the work. And

so, they break up a complete task into component parts, and they hand out each piece of work to a specialist, i.e., an

associate or paralegal, and they're all doing that task running in parallel. And

the keyword there is parallel. They're

not waiting for one person to finish before the other one starts, they're doing it all at the same time. And

that's a really important distinction.

So, sub-agents, when we're talking about the Claude ecosystem, sub-agents don't talk to each other. They take their work from the manager agent, they do their work, and they report back to the orchestrator or the manager agent. The

orchestrator is what pulls it all together. Now, that's different from

together. Now, that's different from agent teams, and this is a key concept to understand the difference cuz we're going to talk about both in this video.

But agent teams is where the agent can actually communicate with each other.

So, there's still a manager agent, but there might be five sub-agents who are working on different tasks. Instead of

reporting back up to the manager agent, they can talk to each other and get updates on how each are going and what's the progress and whether there's dependencies they can start to work if they were waiting on a dependency to be finished first. And that's a more

finished first. And that's a more advanced pattern, and I'll cover it briefly, but for most legal workflows, sub-agents are really what you need and where you should be focused.

Quick note, before we get into the workflows, this is exactly what we build for firms here at Sidebar X if custom sub-agents pipelines for contract review, due diligence, matter intake, the kind of work that I'm about to show you. Not only does it make your team

you. Not only does it make your team more productive, it makes them more productive than they would be using just a single agent. Because a single agent running five tasks in sequence is one x productivity. The same person running

productivity. The same person running five sub-agents in parallel is completely different productivity. We

also run workshops to coach legal teams to use sub-agents or sub-agent teams in their day-to-day. And so, it's not

their day-to-day. And so, it's not generic AI training, it's custom-built and customized to your specific firm or legal team and the way that you work. If

either of those sound interesting or relevant, I'll leave a link in the description below.

Start by asking yourself three questions. And really, these are the

questions. And really, these are the only three questions you need to ask yourself. Number one, can this task be

yourself. Number one, can this task be broken down into independent pieces where piece two doesn't need piece one to finish before it can begin?

Number two, is the volume just too high that it's going to blow one context window and you're not going to be able to do much analysis before you need to start a new context window. So, are you uploading 50 documents at a time? If you

are, then that might situation where you want to use a sub-agent. And three, do different sub-tasks benefit from different reference material or a different focus? If you answer yes to

different focus? If you answer yes to any of those, then you're really looking at a sub-agent workflow.

So, let's look at a couple of real-life examples, right? So, let's start with

examples, right? So, let's start with M&A due diligence. So, you've got Let's say you've got 200 documents in a data room. You can either do it with one

room. You can either do it with one agent that reviews them sequentially, or you have five agents that has one document category per agent. So, IP,

employment regulatory litigation whatever it might be, and you run them simultaneously. Now, the orchestrator

simultaneously. Now, the orchestrator synthesizes that into a matrix. And the

thing is, that's not actually just five times faster, it's even faster than that because the orchestrator, i.e., the

single agent, only has to read the summaries, not the raw document. Let's

look at multi-jurisdiction research. So,

let's say a client operates across 12 states and the EU. They spawn one sub-agent per jurisdiction. Each returns

the relevant statute, case law, and pending changes. Orchestrator

pending changes. Orchestrator cross-references the conflicts. Instead

of 13 sequential searches one after the other, you get 13 simultaneous searches.

Contract review pipeline. A really

common workflow that we build automations and other skills for inside of legal teams. So, let's look at a scenario here, right? So, sub-agent one checks against your firm's playbooks.

Sub-agent two compares against your last five similar deals. Sub-agent three

flags regulatory compliance issues. The

orchestrator compiles the redline and risk summary. Three parallel lenses

risk summary. Three parallel lenses instead of one agent switching context three times. And when an agent is

three times. And when an agent is switching context and it's got a lot of different information it's got to ingest that are not necessarily focused and similar, then you're going to get drift in in terms of the output and in terms

of the quality as well. And then

finally, matter intake and conflict check. So, three different tasks. You've

check. So, three different tasks. You've

got conflict check, issue spotting, and party research, opposing party research.

They have completely different modes.

One is a database lookup, one is analytical, one is search-intensive, right? And so, there's no reason to do

right? And so, there's no reason to do them in sequence. They are perfectly suited to parallel sub-agents. The

orchestrator aggregates the intake memo.

Okay, so we've talked about what type of task are good for running sub-agents. Go

back and look at those three questions cuz an answer yes to any one of those is a good candidate, as I said. But I think it's also good just to re-emphasize and hammer home when not to use sub-agents.

And so, don't use sub-agents for sequential work. So, drafting a

sequential work. So, drafting a settlement response after reviewing the opposing party's demands. That's step

one, step two. You can't run them in parallel, obviously. A sub-agent would

parallel, obviously. A sub-agent would just add overhead, complexity, and there would be zero additional benefit.

And here's another one that might not seem that obvious because you might just think, "Oh, AI output is all the same."

But it's actually not. Now, if you're drafting something or writing anything that needs one unified consistent voice, sub-agents may not be appropriate, particularly if it's client-facing. So,

five agents each writing a section of a brief potentially gives you five writing styles. Now, you can overcome that with

styles. Now, you can overcome that with really good prompting, but what I'm teaching you here today are purposely entry-level skills. And so, it's beyond

entry-level skills. And so, it's beyond the scope of this video. And so, just know that if you have got five different writing styles, the orchestrator then has to rewrite the whole thing for consistency. And so then, potentially,

consistency. And so then, potentially, you've done the opposite thing of saving time. You've actually maybe taken about

time. You've actually maybe taken about the same time or maybe even a little bit more time.

Now, definitely don't use sub-agents when you're editing a singular document.

They're going to override each other the same way that two associates would if they're both working on a cloud Word document at the same time.

And certainly don't use sub-agents for privilege determinations. This one

privilege determinations. This one really matters, right? So, just think about it this way. Privilege analysis

obviously requires a unified judgment across all facts at once. And if you're splitting it across agents, then there's no single agent that has the full picture. And I guess a waiver that you

picture. And I guess a waiver that you caused by mistake certainly isn't recoverable. And so, single agent human

recoverable. And so, single agent human review. And my last thing I'll say about

review. And my last thing I'll say about this is, don't over-engineer simple tasks. I see lawyers who learn how to

tasks. I see lawyers who learn how to use sub-agents want to default to sub-agents all the time because they've seen the productivity gains. But like I said earlier, if task is not suited to it, you may end up taking more time than

you actually would just using a single agent. So, if a single agent can answer

agent. So, if a single agent can answer it in two to three prompts, then don't spin up an orchestrator. The complexity

really has to earn its place. Earlier, I

touched on agent teams. We've just spent some time discussing sub-agents, so let's talk about agent teams. I know it can be confusing cuz the names are obviously similar. Sub-agents don't talk

obviously similar. Sub-agents don't talk to each other. Agent teams talk to each other. So, if sub-agents are the senior

other. So, if sub-agents are the senior associate handing work to junior associates who each do their piece independently, agent teams are more like a working group where the associates can actually challenge each other's

conclusions. Just imagine them sitting

conclusions. Just imagine them sitting in a meeting room together.

And I'm not speaking from a lawyer standpoint, but speaking from a standpoint from lawyers informing me, the legal workflow where this tends to matter the most litigation strategy. So,

you want to stress test your theory of the case, spawn three to five teammates, each assigned to argue a different theory or take opposing counsel's position. They debate directly trying to

position. They debate directly trying to disprove each other. The surviving

theory is often the most defensible one.

LexisNexis Prodigy actually runs this in production. A lead orchestrator plus

production. A lead orchestrator plus research, web research, and document agents that coordinate with each other.

Agent teams are more advanced pattern, but they're still emerging in terms of tooling. Sub-agents in Claude Co-work

tooling. Sub-agents in Claude Co-work really are where you should start, and it's where I'm training teams at the moment to focus their time and attention. Okay, I have done far too

attention. Okay, I have done far too much yapping at the expense of showing you actually how to use sub-agents, so let's rectify that right now. So, I'm

here in Claude desktop app, specifically Claude co-work. If you want to know how

Claude co-work. If you want to know how to use co-work, I've made some videos, I'll link them in the description, but I've created a project inside of Claude co-work. Okay, so sub agent demo. If you

co-work. Okay, so sub agent demo. If you

look over here, there's some files in the directory that I'm working out of.

You've got a vendor agreement, and we've got in our firm's contract playbook. And

so, I've pasted in this long prompt here, and you'll see that it says, "You are a contract review orchestrator. You

will review the vendor agreement uploaded to the project using three independent work streams in sequence.

Complete each work stream fully before moving to the next." And then you can see I've defined what sub agent one, two, and three do, and then even given the orchestrator instructions to synthesize all of their output. Now, I

deliberately showed you how you can do this in very, I guess, a formulaic way, but Claude is also smart. If you just give it a prompt and just say, "Work on this in parallel," then Claude will spin up sub agents automatically. You can

also invoke sub agents by creating an instruction in a Claude MD file, which can be specific to the project you're working in, or it could be a company level setting that a specific sub agent

is spun up at the start of every session, or when a specific command is typed in. They can even be invoked

typed in. They can even be invoked through a skill, and a skill is just an instruction for completing a specific process. So, if it's contract review,

process. So, if it's contract review, you can build in parallel agents into that skill to always be invoked when doing specific types of contract review.

Okay.

So, as long as you understand that, there's many different ways that you can invoke a sub agent, then you're going to be able to use them effectively, combined with your knowledge of what type of work is well suited to using sub agents with.

So, let's hit let's go. Get into the habit of reading what Claude's doing.

You can also expand and see more detail.

There's not much detail in this instance, but sometimes there'll be loads of detail hidden behind the expanded menu, which starts to teach you what Claude's doing, and gives you confidence in a lot of instances that it's working on the right things, and

it's doing what you asked it to do.

So, just while that's working, let's just quickly go back and look at what the sub agents are doing, right? So, sub

agent one, "Review every clause against our standard playbook position, flag any deviation." Sub agent two, "Historical

deviation." Sub agent two, "Historical comparison. Compare the key commercial

comparison. Compare the key commercial terms, fees, liability cap, termination, IP ownership against market standard for SaaS vendor agreements with professional services firm." Sub agent three,

services firm." Sub agent three, "Regulatory compliance. Flag any clauses

"Regulatory compliance. Flag any clauses that may conflict with A, attorney-client privilege requirements, B, legal data retention obligations, C, bar association ethics rules, and D, any applicable data protection rules." And

so, very distinct work tracks that can be done in parallel. And this type of workflow perfectly suited to sub agents.

I'm just drawing your attention to the right-hand side of the screen. This is

your progress bar, right? So, we can see the process that it's going to go through, and we can see where it's up to, as indicated with the the blue around the number. Okay, so now I've got an update. All three work streams are

an update. All three work streams are complete.

Uh the agent is telling me the orchestrator agent is telling me that it now has all three analysis complete.

It's going to synthesize its findings and generate the final deliverable.

Just before it's finished its output, I just want you to picture the alternative workflow where you're doing what sub agent one did, sub agent two did, then sub agent three, and then you're pulling

that all together, asking an agent to synthesize it, and come back with a summary, or come back with findings. And

so, just looking at the steps, you can see how long that would take, and how much more productive and efficient you are spinning up parallel sub agents.

And there's the synthesis. One

structured prompt, three complete work streams. So, let's just scroll down.

Let's look at the output really quickly.

Top three must fix items, which is also highlighted here in the output in line.

So, we've got the flagging, the high, medium, low using color coding.

We've got the historical comparison, the regulatory work stream.

And that's the whole review, and under about Well, honestly, it took about, I'd say, 5 minutes. If you were using Sonnet 4.6, which would have been perfectly suited to this type of analysis, it probably would have been done in about 2 minutes.

And so, that's the pattern, right? Swap

out this work stream for your due diligence categories, your jurisdictions, your intake tasks. The

architecture, the way that you prompt Claude, is exactly the same.

And finally, if you want to map your own workflows to this system, I built a free tool called the legal agent workflow planner. You put in your legal task,

planner. You put in your legal task, answer four questions, and it tells you whether to run it as a single agent, sub agent, or an agent team, and how to break it down. The link is in the description. If you want us to build

description. If you want us to build this for your firm, or coach your employees, or your team how to use Claude, how to use agents, sub agents, then the link is in the description for that as well. If you liked what you saw

today, love a subscribe, or a comment. I

publish once a week, and I'll keep doing it as long as it's useful.