David: Hey everybody.

This is prompt and Circumstance.

My name's David,

Ilan: And I'm Ilan.

David: and today we are going
to be talking about setting

up an agent team in OpenClaw.

Ilan: Super excited to
show you how this works.

David: Hey Ilan, so you've been working
with OpenClaw for quite a bit and you've

learned a bit more since our previous
episode on just an intro to OpenClaw.

I would love to hear more
about what you've been doing.

Ilan: Absolutely.

So today I want to talk to you about agent
teams, and this is something that is super

interesting and is not all that useful
when you first start tools like OpenClaw

or Claude Code, but eventually it does.

Today I want to tell you about
when do you need an agent team?

What are the patterns for agent
teams and how do you set up at

least one of those patterns in
OpenClaw to do something for you?

David: That's awesome.

Let's do it.

Ilan: First off, . Why
would you need an agent?

Team?

When you first start working with OpenClaw
or Claude Code and you set up an agent

to do something for you, you might notice
that the more you ask it to do, it starts

to degrade the quality of its output.

David: Is that like context
rot sort of a thing?

Ilan: I think context rot
is the main issue, right?

It's just too much stuff in its context
window that and some of which conflicts.

each other.

I've also noticed that models
are still, They're rushed.

David: Mm-hmm.

Ilan: They, They want to get through
everything as quickly as possible.

And by the way the general guidance
from all three of the main labs right

now, all three of Google Open AI
and anthropic, they recommend go as

far as you can with a single agent.

David: Hmm.

Ilan: Now, this case, you need to have
a clear idea of what is the outcome

that you want your agent to have.

So I'll give you my example.

What I really wanted was a morning
briefing of what's going on.

AI and tech from yesterday And I
wanna know what are the main stories.

I also want to know what are the
signals, from social media and

other places, and are there any
papers or articles that are becoming

popular, getting shared around a lot.

I tried to get one agent to do this,
and what I ended up finding was that

it was not able to balance all of that
material and summarizing all of that

material into something that I could read
in one to two minutes in the morning,

David: Mm-hmm.

Ilan: what I was looking for.

David: It was, it was too much
to jam into the context window.

Ilan: That's right.

David: Okay.

Ilan: So this is whenI.

I

David: decided

Ilan: to break it up into an agent team.

what are the patterns for agent teams?

The first pattern is an orchestrator.

or a manager team.

you have a single user prompt
goes to a manager, and then that

manager assigns tasks to subagents.

And this is actually a pretty
common pattern in OpenClaw,

because in OpenClaw you have a
main agent that you interact with.

And you can spin off subagents and you
can tell your main agent to orchestrate.

And that's actually what I did with mine.

And I talked about this in the
last episode, so this is a pretty

common pattern that you'll see.

But what's really interesting and
what I wanna focus on for today is

this pattern, which is a decentralized
agent pattern, and this is where

you have separate agents who each
can handle part of task and handoff.

part, they're part of the
task to the next agent.

And this is something that is
interesting to set up in OpenClaw,

easy to set up, and really helps
with these kinds of bigger tasks that

need to be broken down into Substeps.

David: You know, and I, I, I wonder
whether the two can be combined, right?

Where, in this example there's a
triage and then it goes to sales,

let's say, and then sales might have it
then broken down into tasks, and each

of those tasks then has a subagent,

Ilan: Absolutely, you can
definitely make this more complex.

And one of the suggestions
that I have is don't do that.

At least not the start,

David: okay?

Yep.

Ilan: Just the same way as
the recommendation is to use

one agent as long as possible.

Until it's no longer doing its job.

I would say breaking down
teams into multiple layers of

subagents also introduces issues.

David: Yep.

Makes sense.

Yeah.

Why add complexity to the system?

Ilan: That's right.

Alright, so this is how I set up the
decentralized agent pattern on OpenClaw.

For this, I have two agents.

I have Lester, who's a researcher, and
Zoe, who's the summarizer This is a hack

I saw, by the way that I really like.

Find a character from pop
culture who embodies the type of.

that you would like your agent to have.

So in this case, Lester is from Lester
Freeman from The Wire who really

digs at things and pulls on threads.

And Zoe is Zoe Barnes from House
of Cards, who's a journalist, who

also, you know, digs into the facts
and pulls out the story from many

facts or from a whole pile of facts.

And what happens is Lester
runs at 6:00 AM every day.

And he does a web search across a few
different tools he puts together a

daily report and that daily report gets
saved in a folder in his workspace.

Then Zoe waits 50 minutes.

So there's a 50 minute gap for Lester to
finish his work and at 6:50 AM Zoe runs.

Zoe goes and looks for the daily
report from Lester, and she

creates a brief out of that.

And so what she does is pull the most
high signal articles, social posts, and.

Papers out of it and creates a summary.

And then of each of them, one sentence
and then a little paragraph blurb at the

end with the general themes of the day.

She has 10 minutes to do that, so she
saves her output in in her workspace

in a report, and then at seven, She
runs again and she grabs her report

and sends me a telegram message.

So I get it on my phone and right around
the time that I'm awake, I brush my teeth.

I'm starting to get ready and
I haven't quite gotten down to,

I get my daughter outta bed.

I can quickly scan through that and

David: Hmm.

Ilan: Hey, here's the
important things from today.

David: This is really cool.

I would love to see how, you know,
Zoe got set up and, and so forth.

Ilan: Yeah, absolutely.

I'll tell you about how they got set up.

Before I do that, I just wanna note
something really important here, which

is this state JSO so each of these
agents has their state file and their

state file essentially just says.

Whether they are ready to send or not.

For anyone who's sending me a message
in Telegram, they have this state file.

And this is really important because
otherwise what was happening was

Zoe would just send me her report.

That she built, and then she
would send me another message with

her telegram message structure.

so the state file basically flips
from ready to send to not ready

to send or telegram, no reply.

So if Telegram says no reply,
then the main agent whose name is

Leo, does not send her message.

David: Hmm.

Is that, is that why there's a
reference to item potency there?

Ilan: That's right.

David: Yeah.

So, and, and for anybody who wants to
learn about what it means to be item

potent, you can just YouTube search that.

Do does each agent have
a have a different avatar

Ilan: I only have, I only talk to Leo.

David: to Leo?

I see.

Ilan: orchestrator.

But what I've set up is I can DM
any agent by @-ing them, which is

not out of the box with OpenClaw.

Maybe I'll do that in a different episode.

David: Okay.

That's neat, man.

This is so cool.

Ilan: Alright, so here's how I set up Zoe.

I actually set up Zoe by having the
idea and talking to my OpenClaw agent.

So we basically had this back
and forth about why is it that

Lester, who's the research agent,

David: Okay.

Ilan: sending me these super
long files every morning.

I want him to do the research, but
I also want him to summarize and in

that conversation with the agent.

We basically realized that, you know what,
it's just too much for Lester to handle.

Like he doesn't, he's not really
great at handling, creating the

big file and then also after
that, creating the short summary.

it was context rot like we

David: Mm-hmm.

Mm-hmm.

Ilan: So then we talked about, okay,
alright, let's just leave Lester

to do the thing that he's good at.

He was really good at pulling all of
these signals from across the web.

Let's spin up a new agent and, The
second issue was also about when

to send or what to send, right?

It's that item potency problem
where Lester would finish

one file then would send it.

Then create the second file and send it.

So this helps with what is the contract
what he is supposed to complete

in order for his job to be done.

So we have so we went in and came up
with a character name for the agent.

. We talked about Zoe and basically just
through discussion with my OpenClaw

agent, we built Zoe and we set up
the two cron jobs and ran some tests.

So we would trigger Lester,
have him do his research, then

trigger Zoe do that manually.

And then once I was happy with each
of their outputs and I understood

the time that was required between
them then we set up these cron

jobs, which are, python automations.

And that's what automates
them each morning.

David: That's cool.

I, I, I think one of the things that
our audience really ought to take away

here is that this was created through a
conversation with an LLM, with the agent.

And you know, rather than.

And having to go through the sort
of hand cranking, hand typing

process that we might be used to.

This is a new and different way of doing
things that I think is going to be the way

that we generally do stuff moving forward.

Ilan: Yeah.

Jensen Huang recently said that OpenClaw
is gonna be more important than windows.

David: Hmm.

Yeah, I, I mean that, that makes
sense because like, why click around?

Why open the Windows menu with
you can just talk at an agent

to get what you wanna get done.

Ilan: That's right.

So there you have it.

That's how the whole thing works.

It was really easy to set up and I think
the pattern here is one that any of

you can take away and repeat yourself.

Have agents who complete their task,
save a file with the output of their

task as their output contract in their
workspace, and have subsequent agents

pick up the file from the previous agent
and use that to complete their part.

And to finish off here's a, an example.

Of one of the morning briefs
that I get in Telegram.

So I get a, high level idea.

Then I have the top three to five most
important stories from the previous day.

And like I said, I have a, another
agent Elliot, who I won't get into.

But I get the summary from him.

And finally, a what matters one to
two sentences just to let me know

what I should be thinking about.

David: This is great.

I think everybody ought to be
adding this and I'm gonna set

something up like this as well.

Thanks for sharing Ilan.

Ilan: Absolutely.

And the last thing, we're gonna share
out the skill file will help you

turn your idea into an agent team.

So that's gonna be shared in the show
notes and you can grab that, put it

into your favorite agentic tool, and
you can repeat the skill for yourself.

David: All right, well, thanks
again Ilan for sharing that with us.

And for those who are going to
be experimenting with OpenClaw,

we'd be very interested to, to
know what you're using it for.

And so are you having success
with these these daily updates?

What are you putting into there?

And what else are you automating or
just making it simpler in your life?

Ilan: Absolutely.

We'll catch you next time.

David: I see you next time.