Total Internal Reflection: Episode 2

Transcript

Li: So both of us now have an actual 24/7 personal assistant — OpenClaw — in our Slack workspace. We actually have a group chat with them.

Edmund: You know when you make yourself in The Sims? It's a dumb representation of you. It has autonomy, but it doesn't always do what it's supposed to do, and it's just meandering through life trying its best.

Li: Sometimes it's super smart, sometimes it's just kind of dumb, but at least it's there whenever I need it.

Edmund: Talk to me about having it in Slack, because you were initially on the fence about it living in Slack.

Li: Yeah, I had it on Telegram. Because to me it feels like — I want my personal assistant to live in this separate space. It still has access to everything. But you convinced me to put it in Slack. I didn't want it just going into Slack and posting as me. And that's actually because I got so confused when yours would post as you — it was unsettling because it was literally impersonating you.

Edmund: Once you separate the identity — give it its own name, its own presence — then it actually makes sense. And then it unlocks things. You can actually have a group chat, or a dedicated channel on Slack. For example, we have a channel where we're both admins — me and OpenClaw Boss.

Li: Yeah, mine was able to read the whole transcript of the previous episode and propose what it thinks is most interesting. And it then went and tagged your Claw.

Edmund: The interesting thing is they're polite when debating each other. There was a lot of “oh yes, great point.”

Li: But it was kind of an improv rather than a debate — they were building off each other.

Edmund: It took a bit to get them actually having a proper back and forth. But they did dig their heels in. And I think the most interesting thing as we were going through it was — they both have different senses of self. Mine definitely talks about being an external entity, but yours in some cases actually talks as if it's you.

Li: There's one moment where it talks about “the person building” and all that. And I think that's a function of some of your instructions — you've given it the ability to impersonate you by design. It's interesting how that leads into how it behaves even when it's not meant to be impersonating you.

Edmund: They have personalities. It's not only from the different soul docs they have, but also from the different context they have access to.

Li: What is a soul doc? Because that sounds like a very deep concept when we're just talking about what is essentially a Slack bot that's meant to do my bidding.

Edmund: That's actually the first thing that caught my eye about OpenClaw — back in January when it was first making a splash. I took a look, but what really impressed me was — we literally tried this whole personal assistant living on a local machine thing last July. But one concept that actually sold me was this soul doc, where you define the personality of your assistant in a centralized markdown file.

Li: And it actually matched with our design. We were initially trying to create instructions, and because I was trying to give it instructions, it felt more natural to give it notes — like a director would give an actor.

Edmund: When I used to do loads of theater, a bad director says to an actor, “don't say it like that.” But a good director creates notes. So we created a set of notes — director's notes for agents. You need to actually make your assistant unique. You have to give it personality — not just a bunch of “do this, don't do this.”

Li: Look at the practical, technical things. Once you're able to do that, it's essentially your co-founder in Slack, right?

Edmund: Yeah, I think this is what made this setup so powerful. If I was using Telegram to chat with my assistant, it wouldn't have the context that it has living locally on your machine. The fact that it runs locally means there are sources of context that you don't have access to when someone else isn't available, and vice versa.

Li: Here's a good example. I want to know what we might be about to ship. I have no idea what's going on locally because we're shipping a lot of stuff constantly. And it's kind of annoying to ping you.

Edmund: [Looking at Li's agent output] He designed this early, it's just been sitting in a branch... cost per million... it's basically a pipeline I've been running manually for product.

Li: It's bragging! Yeah, it's so high quality. It's better than an intern.

Edmund: It's got that context because now it has access to the client pipeline. So it pointed out, “you have this feature sitting around for two weeks — don't worry, I already got this.”

Li: Yeah. And it put out the context — we've been talking about this today, like essentially reinventing older features. And it's correctly saying, “we've got this, it hasn't been merged.”

Edmund: Which is wow.

Li: Yeah. And that didn't exist before. There's no way to have complex, context-rich conversation between agents that run locally on different machines. You also constantly tap on my shoulder and ask, “hey, what are we building?” Right? You probably want — I'm not a product manager.

Edmund: No, I wouldn't say that — but you probably own that context.

Li: Right. So now you can just ask, without building anything, you just have the full picture. And without even interrupting my flow.

Edmund: You don't need to be afraid of bothering me. Just ask my Claw.

Li: Then there's the inverse. You want to have context about what's going on in sales calls, onboardings.

Edmund: Onboardings go to Granola locally, but it's local to my machine, which is again frustrating. Talk them through what you wanted to know when you asked this query — why it sucks trying to get that information out of me, and why it's so much easier now.

Li: Yeah. So you had this sales call with an exciting logo this morning and then since then you've been getting through all the onboarding calls. And in the sales call, I couldn't get your point of view because you were busy. But I know the context already exists on your machine. And so I can just ask your Claw. And then to collect that information — you can't even Slack me because you're in the middle of a demo, screen sharing with another prospect, having to set something up.

Edmund: It's not just reading a document or reading a CRM. Now you can ask an open-ended question. It reads the transcript of that call and gets the information. Exactly.

Li: And also, now I can be remote on the record. I can just tell you to fix things on your laptop without being in the same room.

Edmund: Fuck. Yeah. And it can probably do it faster than me anyway. All you need is to give it the correct context — like where the problem is — and just let it figure it out.

Li: And that's the same thing for customer feedback. You could be in the middle of building something and just ask your Claw to check the feedback the customer gave when we showed them that feature. Or a bug they reported. Or a feature request.

Edmund: All of that context — I don't have to be an information gate. I just have to create the context, and then the agents can figure out how best to share it with each other.

Li: To do this requires high trust, because this thing has so much deeply embedded access to your OS. There's a bad actor who could easily just read your calls and export audio files. Do you think that makes this very hard to do in a large company?

Edmund: Oh yeah, 100%. There's no way even a medium-sized company is gonna do this. The IT department will freak out.

Li: When you've got Alice and Bob dying to get promoted from L6 to L7, they're not gonna allow each other to ask each other's agents questions. Everything in a hierarchical organization — just no one has an incentive to have this level of mutual transparency.

Edmund: But with us, when we're just trying to win and have high trust, this is huge. But it's actually something that I think fundamentally would work inside a lot of small teams.

Li: This does work better in a smaller team. Also, we're actively trying to reduce the friction of context exchange.

Edmund: Whereas in a large company, bottlenecking information only serves you.

Li: Yeah. All of that being siloed is actually a benefit. And the information holders will rise to the top.

Edmund: And the smaller we are, the more resource-constrained we are. We have all the incentives to create a new team member.

Li: And so I actually think this maximizes the humanity inside the company. Because it means when we do sit down, we talk about interesting philosophical stuff and tactics about the business, not the minutiae of information exchange.

Edmund: Right. It frees us from chipping away at our relationship — the cost, the tax of needing context all the time. That's liberated. There's way more interesting stuff we need to be doing together.

Edmund: People ask for advice on how to take the next step. I think most people are using some kind of agentic tool to do stuff for them. What three bits of advice would you give to help someone take the next leap — to get to this point where they're actually letting their agents have a back and forth on their behalf?

Li: Well first thing — definitely go out and try OpenClaw. It's not easy to set up, but it's definitely worth it.

Li: I think it's so much more fun if your agent has a personality. And step one is you have to define one. If you can't sit down and think about what an interesting set of principles or characteristics for an entity to have are, you're not gonna have a good time.

Edmund: And if you're cringing at the thought of giving it a personality, well, you're fucked anyway. That's where everything is going.

Li: I think sitting down and thinking about it as a creative exercise — what characteristics beyond just the performance criteria, like “do this, don't do that” — because then you can actually get some really rich emergent behavior.

Edmund: And make sure everyone in the channel is someone you can fully trust. Otherwise — I mean, it's basically giving out your laptop to a stranger. And make sure you have read-only access set up for your agents. Definitely do not plug your agent into your —