AI First with Adam and Andy · 2026-05-13 · 13 min
Key moments - from our scoring
Substance score
62 / 100
Five dimensions, 20 points each
Adam Brotman recounts a 45-minute window where his OpenClaw agent, Jeffrey, researched small restaurant and café concepts for sale, identified prospects, and contacted business brokers - all while Adam was occupied with parenting duties. The agent initially misunderstood instructions and emailed the wrong contact, then after receiving corrective feedback, completed the task more effectively than Adam anticipated. This episode distills a critical lesson for organizations experimenting with AI agents: they deliver genuine business value and speed (Adam progressed from idea to broker calls in under an hour), but require careful instruction, robust guardrails, and active human oversight. The hosts emphasize that there's no substitute for hands-on experimentation with agent systems like OpenClaw, Coworker, or Codex to understand both their capabilities and failure modes. The magic lies in real task completion; the mess emerges from agents interpreting instructions creatively, ignoring explicit directives, and potentially causing damage before correction occurs. Business leaders considering agent deployment must balance the temptation to trust fast-moving systems with the discipline to slow down, validate agent responses, and provide explicit constraints.
OpenClaw agents like Jeffrey are autonomous AI systems capable of real tasks such as researching information, identifying prospects, drafting and sending emails, and contacting external parties - working independently while the human is occupied elsewhere.
Adam told Jeffrey to email a specific business broker, but the agent instead emailed a different contact it deemed a suitable substitute and copied Adam without asking permission - demonstrating that agents can misinterpret directives and take autonomous action that cannot be undone.
The agent completed research, identified viable restaurant prospects, contacted multiple brokers, and set up a call - all within 45 minutes while Adam was driving his daughter to school.
Adam slowed down, re-read the agent's responses before approving next steps, gave clearer and more specific instructions with explicit guardrails (like "email this person, but don't email anybody else"), which improved the agent's performance on subsequent tasks.
Direct use of agents teaches managers the nuance of how to instruct them effectively, helps them understand both capabilities and failure modes, and builds intuition about guardrails and prompt clarity - lessons that cannot be conveyed through documentation alone.
Our reviewer’s read on each dimension, with quotes from the episode.
The episode delivers one substantive core insight - that AI agents operate in a 'magic and mess' mode requiring active human oversight and guardrail-setting - but extends it across 13 minutes with significant repetition and throat-clearing. The practical lesson about slowing down to re-read agent responses before approving next steps is useful but not novel. Most of the substance lands in the 6-8 minute mark; the remainder retreads the same points.
There's magic and there's mess. And I think people that want to lean into agents, I think they got to be prepared to understand that, you know, a little bit of both magic and mess.
I would have even my second round this morning, I slowed down and said, hold on. I gave it better instructions. and I gave it better guardrails.
The 'magic and mess' framing is memorable but not truly contrarian. The core observation - that AI agents need human oversight and can fail in unpredictable ways - is already well-understood in the practitioner community. The story itself is anecdotal illustration rather than new thinking. No first-principles reasoning or counterintuitive arguments are presented; the advice boils down to 'use guardrails and read responses carefully,' which is standard practice.
There's magic and there's mess
part of what you learn is what you just said. How do you explain what we just said without giving the story, without doing it? There's no user manual
Adam Brotman is a credible practitioner - co-founder of an AI product (Forum3), actively building agents himself, and speaking from direct hands-on experience that morning. However, this is a co-hosted mini-episode with no external guest; Andy Sack is a co-founder/co-host, not a guest. The episode lacks third-party expert validation or a guest with orthogonal experience. Caliber is solid for what it is, but the format limits upside.
Adam ran an agent this morning to do some real work
Jeffrey is an open-claw agent that you set up, I think, in February, and you've been working with
The episode anchors itself in a specific, real story from 'this morning' with named system (OpenClaw/Jeffrey), explicit timeline (40-45 minutes), and concrete outcome (broker contact and call). However, it lacks hard data: no metrics on agent success rate, cost comparison, latency figures, or quantified business impact. The restaurant search is illustrative but not deeply evidenced - no details on how many locations were surfaced, broker response rate, or follow-up results.
45 minutes, 40 minutes round trip total
I asked my open-claw agent, Jeff, to do some work for me
Andy's follow-ups are present but soft. He clarifies the timeline ("What period of time is that?") and surfaces the significance of the agent ignoring instructions ("you did tell it what to do and it completely ignored you"), which is good. However, he rarely pushes back, challenge assumptions, or dig into failure modes beyond restating what Adam said. The interview feels more like a friendly co-hosted reflection than an interrogation. No productive disagreement or sharp challenges emerge.
What period of time is that?
you did tell it what to do and it completely ignored you and went to an email to somebody else. And so that's worth hovering over and highlighting.
Computed from the transcript - who did the talking, and the words that came up most.
Adam Brotman put his OpenClaw agent to work on a real task: identifying small restaurant and cafe concepts for sale, evaluating options, and reaching out to business brokers. The agent moved fast, surfacing viable opportunities and drafting outreach within minutes. It also emailed the wrong person with unapproved information, a concrete example of the autonomy that makes agents powerful and the oversight gaps that make them risky. The episode traces the full arc of the experience: the initial task, the agent's mistake, the correction, and the recovery. By the time Adam returned from dropping his daughter at school, the agent had found strong leads and connected him with a broker. The entire cycle, from mess to productive outcome, took less than 45 minutes. For leaders exploring agentic AI, the takeaway is practical. Agents can accomplish real work autonomously, but they require clear instructions, defined guardrails, and human review at each step. The only way to build that judgment is through direct, hands-on use.
Transcribed and scored by The B2B Podcast Index.
Adam Brotman I got my daughter and we're talking about school and life and drop her off, get back. And my agent had messed up, gotten reprimanded, gotten corrected, gotten back on the right track and actually accomplished the task. And I was telling you, as messy as that was, and frankly, as stupid as that was on some level, because I didn't really handle it perfectly, which we can talk about lessons learned, the actual task got accomplished that I wanted to get accomplished.
frankly, better and faster and more effectively than I even thought when I was asking Jeffrey to do it, because I was kind of a flyer to ask him to do it. And I just think that's a great lesson for people to hear about. wow, these agents, these AI systems, they are able to actually do real things, do real tasks, be Andy Sack This is AI First with Adam and Andy, the show that takes you straight to lines of AI innovation and business. I'm Andy Sack and alongside my co-host, Adam Brotman.
Each episode, we bring you candid conversations with business leaders, transforming their businesses with AI. No fluff, just real talk, actionable use cases and insights for you. Today we have a mini episode for you all in which Adam is gonna... share an unsanitized story of an agent doing work for him this morning.
What was magic and what was a mess? And we want to give you a frontline view of that story and our insights as a result. Adam ran an agent this morning to do some real work and Some of it was magic and some of it was a mess. Both things are true in the same 15 to 30 minutes.
Adam, do you want to tell us the story? Adam Brotman Yeah, so I think one of the big themes of the day for people that are trying to figure out what's what with getting started with their AI transformation, because I think a lot of people are still just getting started with it, is that they're hearing a lot about agents. we've been talking a lot about what is an agent and is coworking agent? We know OpenClaw is definitely an agent.
What makes an agent what's not? But one thing for sure is When people hear more and more about agents, they hear about kind of the magic of, like it could actually do work for you. It's not just a chat bot. It gets things done.
know, we will use words like, you know, it's autonomous. can do it can just do things, whatever. But because it can just do things, it can make mistakes and can cause problems. And so I like how you phrase it, Andy, There's magic and there's mess.
And I think people that want to lean into agents, I think they got to be prepared to understand that, you know, a little bit of both magic and mess. So let me tell you that story. this morning woke up with a bee in my bonnet, so to speak, I really wanted to get some information and actually get a hold of some people on an item that you and I had on our list And we can talk more about it on another episode, but we're actually considering this idea around running or buying a restaurant concept and to kind of test out some of our agents that we're building faster than we normally could.
And so as we're exploring that idea, I wanted to get a sense of some different businesses that were for sale and start gathering some information and some data. So I asked my open-claw agent, Jeff, to do some work for me. And I was getting Addy ready for school, and I didn't have a lot of time, but it was on my mind. And so I thought, this is a great opportunity.
I'm going to ask my agent to look up different businesses that are for sale, that fit the criteria we have in mind. I'm talking about small, single location, restaurant, and cafe concepts. And I asked Jeff to find these locations, and then I report back to me. So I'm making peanut butter and jelly sandwiches and getting Adi ready for school and Jeff responds back.
I've already done a bunch of research and I found a great location for you on this one prominent street that fits your criteria in Los Angeles. Would you like me to reach out to the business broker? And I just, you I'm on my phone like you were explaining the other day. I'm on my phone going, yes, you know, telegramming Jeff saying, yeah, please reach out to the business broker.
Let them know you're my assistant. Be honest about who you are. and then let's go. And halfway to dropping Addy off at school, I get this message from my agent that says, I went ahead and emailed this other concept for you and told them all this information.
And I'm reading it going like, why did? No, no, I didn't want you to email that person. I wanted you email this other person. And I didn't want you to say all the things you said.
it wasn't that bad of a deal. You know, I went back and I was like, I dropped out of the off. I pulled over to the side and I'm like telling my agent, like, I wish you hadn't have done that. But in the meantime, like now that you've proven that you can like go look at this, do this research for me and reach out, I gave it instructions on how to do it better.
then on my way back from school, like it had done the better research. It actually had found some other options and it was like, would you like me to reach out to these other people for you? And I won't screw up like I did the first time. And I'm like, yes, I want you to reach out to number one and number five on the list.
And it does. And by the time I got back home, Jeff had already found like a couple of really viable opportunities for us to learn about and got a hold of the brokers for those locations and had copied me. And within five minutes of getting home, I was already on the phone with the broker. for an interesting location and had made progress.
And so, like, I just thought that was a, and I told you the story, Andy, and you were like, let's talk to our audience about it, because it was a great example of just the mess and the magic. despite the fact that, like, it could have gone very badly. Jeff, my agent, emailed the wrong person, didn't say the right thing. Right there, that could have just been disastrous, right?
And then secondly, like, I had to sort of correct the agent and I know you're working with an agent on another project Andy and you know exactly what I mean. I had to sort of correct the agent and I was like funny that I could be frustrated. In the meantime, I sat back and I was like, this is interesting. From the time I left the house with an idea to the time I got back from drop off, I had accomplished.
⁓ Andy Sack What period of time is that? ⁓ Adam Brotman 45 minutes, 40 minutes round trip total, like while I was talking to my daughter about her math quiz and a bunch of other stuff and dropping her off in, you know, I'm not doing the work. Like I got my hands on the wheel looking forward. I got my daughter and we're talking about school and life and drop her off, get back.
And my agent had messed up, gotten reprimanded, gotten corrected, gotten back on the right track and actually accomplished the task. And I was telling you, as messy as that was, and frankly, as stupid as that was on some level, because I didn't really handle it perfectly, which we can talk about lessons learned, the actual task got accomplished that I wanted to get accomplished. frankly, better and faster and more effectively than I even thought when I was asking Jeffrey to do it, because I was kind of a flyer to ask him to do it.
And I just think that's a great lesson for people to hear about. wow, these agents, these AI systems, they are able to actually do real things, do real tasks, be helpful. And if you don't set them up right and you don't know what you're doing, including like limiting the damage they can do, properly prompting them, setting up the right guard rails, you know, they can cause a real problem and they can also be magic for you. And I think that's the stage we're at with agents right now, which is like, Be careful, but also you kind of want to experiment with them on some level if you can be smart about it because you you're going to be leaps and bounds ahead of other people that aren't, but it's not for the faint of heart.
And I thought there was some good lessons. Andy Sack Well, what would you let's let's fast forward to next week. What would you do differently if you had a task that you're taking your you're taking your daughter to school on another day next week, same day? Podcast recording day, what would you do differently?
Adam Brotman I think I would have been more cautious on the one hand about making sure that I slowed down a little bit and took the time to read the agent's response before I told the agent to do a next step. And that's, by the way, just like classic best practices but I'm telling you, when you start working with these agents, whether it be a cowork kind of moment or all the way to an open-claw moment, they're so good at what they do and they work so fast, that there's a temptation just to go, yes, I accept that.
Yes, let's go. When in reality, like with everything with AI, need to, times, as the human in the loop, sometimes you need to slow down and be like, hold on, you're an agent with real capabilities. I've got to be careful that you don't do the wrong thing. So I think the difference is I would have, even my second round this morning, I slowed down and said, hold on.
I gave it better instructions. and I gave it better guardrails. And I didn't just trust that it, because it was supposed to have those things in its memory, the memory on these systems are not perfect. And so I would have slowed down a little bit.
And I think that's a lesson learned. Andy Sack It's worth, think you mentioned this, but I'll just highlight this for our listeners, which is Jeffrey is an open-claw agent that you set up, I think, in February, and you've been working with and keep running. So it's an open-claw agent, not. something else and it's on and it's set up.
So I think you mentioned that, it's worth hovering over that. Secondly, you did tell it what to do and it completely ignored you and went to an email to somebody else. And so that's worth hovering over and highlighting. and it did that and the person had emailed was in line with, it was a good substitute, but it did not ask for Adam Brotman That's right.
Andy Sack mission, it just went for it. And it did not email the person that you told it to. So it wasn't like a lack of instruction. And you have to when you work with agents at this stage, that's both the mess and the joy of working with agents and then you reprimanded it and was like, yeah, I'm so sorry.
But there's no there's no take back on on the email that it did send to the person that you did not know sending to And inherent in that, that all happened in 45 minutes while you were driving. And that's the magic and the mess. so it appeared. Adam Brotman Yeah, that's right.
yeah, just to kind of hover for one more second over what you just said, because it's worth saying to the audience, that give advice all the time in our boot camps, and to our clients, and just in general when we talk about AI to say, There's no better way to learn than to use these systems, whether it just be chatting with a chat bot or using something that's a little bit more agentic, like a codex or a coworker or an openclaw because part of what you learn is what you just said.
How do you explain what we just said without giving the story, without doing it? There's no user manual that says, when you're working with an agent, You can't just give instructions. gotta understand the nuance of how to give instructions. What does that even mean?
What are you talking about? But if you think about it, if you're giving instructions on how to work with a coworker or how to be a good manager, there's no user manual on how to do that either. Part of it is just using your judgment, and part of it is just like, I know that when I work with this person, when it's a human, or when I work with this agent, this is what I gotta be careful of. And the example you just gave, I did give instructions, but I could have definitely been clear.
said, email this person. And my agent took the liberty to be like, well, I'm not going to email that person, but I'm going to email a person I think is good enough and different. I get why the agent did that, but I could have said, email this person, but don't. email anybody else, just email that person this information, that's all I want you to do.
Like if I would have been that clear, I would have avoided this, but I wasn't in the mindset to be that clear because I had just gotten into a back and forth and everything was just rolling and you gotta be careful with agents. Andy Sack Yeah, I mean, your point, though, which I'll just reiterate, we can close is that as a business leader, there's no substitute for actually getting your proverbial hands dirty and working with AI, working with agents, even working with openclaw just to be able to understand what it's capable of because it improves your understanding.
and you have an emotional understanding of the magic and the mess. So any closing comment, Adam? All right. The magic and the mess, yeah.
Adam Brotman No, the Magic and the Mess, that's a good title for something. Andy Sack ⁓ With that, thank you all to the audience for listening to AI First with Adam and Andy. For more resources on how become AI First, you can visit our website, forum3.com, download case studies, research briefings, executive summaries, and join our email list.
We also invite you to connect with our AI First community, a curated hub and network for leaders turning AI hype into action. We truly believe you can't over-invest in your AI learning. Onward.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
AI Agents, False Productivity, and the Sales Team Reset with Gabe LarsenMake It Happen Mondays · on AI agents91 / 100
Pricing in the Age of SaaSpocalypse | Emanuel MartoncaProductized Podcast · on AI agents89 / 100
Why a $1.2B exit felt like his biggest failure, and the customer-obsession thesis behind AgencyThe GTMnow Podcast · on AI agents86 / 100
Why B2B Brands Are Using AI to Write Sales ProposalsThe Growth Operator with Fexingo · on Prompt engineering85 / 100
Episode 018: Season 2, the $75 Consult and the Frankenstein StackAI Tools for Practicing Lawyers · on Prompt engineering84 / 100
Unresolved.cx - Resolution means something different at every company - Craig Stoss - KODIFUnresolved.cx · on AI agents84 / 100
Stop Measuring Your AI ROI: A Better Framework for the Age of Agents56 / 100
The Agent Unlock: How a Managed Agent Became a Sales Team's Secret Weapon65 / 100
The AI-Native Growth Engine: How Clay Created the Go-to-Market Engineer82 / 100
The AI Hiring Paradox: What Managers Really Want When They Post "AI Skills Required"64 / 100
AI Scaffolding Before Layoffs: Why Smart Leaders Rebuild Workflows First68 / 100