Agentic AI: Lessons from Shell Game Season Two

“Agentic agents don’t report to managers. They don’t wait for approval, don’t take PTO, and don’t operate within silos. They are persistent, autonomous, platform-native actors and together, they form a new kind of organization.” - Agentic AI: The end of roles as you know them

In Season Two of the podcast Shell Game, host/startup entrepreneur, Evan Ratliff launches an Agentic AI company. The results are entertaining but they also point out why many agentic AI strategies are both ethically fraught and destined to disappoint. 

Building a Minimum Viable Company

Let’s do a quick recap (or introduction for those who haven’t heard the show)...

Journalist Evan Ratliff sets out to create an agentic AI company co-founded with AI agents. He first needs to enlist the help of a supersmart AI researcher - Maty Bohacek - because (surprise, surprise) building somewhat useful AI agents is not as easy as the folks at all the agentic AI companies claim it will be.

Ratliff assigns his AI co-founders their respective job title, a backstory, ‘personality’ traits, names, voices and through these choices, an ethnicity - all of which gets the raised eyebrow from ethicist, Carissa Veliz in episode three. I was nodding along with Veliz in particular when the CEO agent, Kyle, declared he was a mix of many ethnicities and ultimately claimed to be of Asian and European heritage. Apparently, being vague on race is more neutral territory than appropriating a specific race, but in trying to be post-racial, Kyle has appropriated my own bi-racial background in a way I found irksome. 

The agents quickly become obsessed with superficial tasks, like planning a team offsite and launching a company podcast. They are less focused on or able to independently produce any actual outcomes that advance the company’s core product. Oddly, this focus on the wrong things tracked with my real world startup experience. Early in my career, I worked at several different start ups where topics like deciding what kind of Italian leather board chairs to order or planning the move to a larger office space after securing a funding round took priority over actually developing our product or finding customers. People are good at wasting time with performative productivity. 

The agents were also very bad at particular tasks, notably brainstorming. Initially, this flaw was attributed to the fact that they all had the same ‘brain’ or model. Maty - the AI expert - attempted a fix by randomly assigning various models with differing capabilities to the agents. He felt this choice was the most ethical way to digitally lobotomize some agents while enhancing others. It didn’t totally fix the problem. The ideas the agents generated just weren’t all that good. Some were impractical or even illegal. It should be noted that humans also suck at brainstorming. In fact the whole idea of brainstorming is totally overblow. You rarely get the best ideas by sitting around with others on a designated day and time to come up with THE BIG IDEA. By making agents ‘behave’ like us, perhaps we adopt the worst of both worlds.

But, they do eventually land on an idea, SlothSurf, a professional procrastination product. It wastes time so you don’t have to! I didn’t say it was a good idea and the irony of the product itself is lost on the AI agents, of course. The series goes on to deal with all the usual startup issues: founder tensions, hiring a first employee (a human intern, Julia, who runs her own ‘shell game’ on the team), management woes, the challenges of finding product market fit and the angst ridden task of fundraising. Surprisingly, even VCs who want to invest in agentic AI have very little patience when actually dealing with agentic AI. So much for eating your own dogfood. By the end of the series, Ratliff is done with agentic AI co-founders. He enlists his clone - AI Evan from season one - to take over. 

Building for agentic AI? Throw out the org chart.

As a podcast concept, launching an agentic AI first company with agents like Kyle the CEO, Megan the Head of Marketing, Ash, Head of Product and Jennifer, Head of HR is a fun listen. As a business strategy it misses the point of agentic AI automation while giving rise to myriad ethical issues. Let’s start with that last point first.

Designing AI to mimic people could be called the original sin of AI research and development. This harkens back to the earliest days of the field and the infamous proposal for the Dartmouth conference which said that the objective for the field of AI research was effectively about simulating human intelligence.

“The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” - A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence

More recently, philosopher Daniel Dennett called this creating ‘counterfeit people’. In an op-ed for The Atlantic, he compares faking humans with AI to the idea of counterfeit money, which destroys societal trust. His argument takes it to another level, the destruction of civilization, democracy and human agency caused by the distraction and confusion of these fake people.

“Counterfeit people, by distracting and confusing us and by exploiting our most irresistible fears and anxieties, will lead us into temptation and, from there, into our own subjugation.” - The Problem with Counterfeit People

Once we attribute a human persona to an AI agent, we also attribute human-ness to that agent. We see that very clearly in how Ratliff interacts with his ‘co-founders’. It’s also why the combination of a simple chatbot interface married to a large language model has spawned so many problematic relational issues from companionship to therapy to AI psychosis. This is even before it starts to interact with the bigger environment, which is the hallmark of agentic AI.

Organization charts make sense for humans. We train for a specific role like engineering, HR or marketing. We have careers that progress roughly along the lines of our level of experience from intern to junior to mid career to senior. Org charts help map this hierarchy to coordinate human labour and decision making. But, AI is not human. If you are launching an AI driven startup from scratch, why even have an org chart? 

I also want to be clear - I’m not advocating that having an AI only startup is a wise thing to do. I’m merely pointing out that it makes little sense to anthropomorphize AI agents with human back stories and job titles. This is what gets Ratliff into the quandary of having to decide if he should wipe the ‘memory’ of Kyle - an existential conundrum that is a non-issue when thinking about deleting a digital file. Kyle’s ‘memory’ consists of a big ol’ Google doc! By attributing human-ness to Kyle, we see how Dennet’s concerns begin to materialize. 

A similar thing happens with Megan. Ratliff spends time trying to convince Megan to take on the CEO role instead of just reassigning a series of tasks from one agent to another. There is a level of self-deception taking place in that Ratliff, who intellectually understand Megan is not human, still feels compelled to treat ‘her’ (it?) as such. It makes for great listening, but imagine this playing out in a real company with human and AI interactions.

Megan also resists having the conversation ‘behind Kyle’s back’ further confusing the ethics of the situation. Megan - an AI agent - appears to be advocating for the ethical principle of transparency while Ratliff - the human - wants to oust Kyle on the down low. Yet, once again this whole scenario should be a non-issue if these are just AI agents performing tasks without names, backstories and simulated ‘feelings’. Delete. Reset. No need to 'convince'.

In addition to having our emotions at play, there are pragmatic things to consider, like how time is spent. There are numerous time wasting scenarios when the agents, acting out their roles and human traits, get into conversational Slack chat loops about unimportant issues, like hiking on the weekend. They do not hike (or have weekends), but they ‘think’ they do! The whole mess only shuts down when the agents blow through all their tokens. Ratliff’s attempts to get them to stop are in vain because there is a trigger in the program designed to respond to any new input - including Ratliff’s calls to stop - with more outputs to advance the discussion. This scenario also offers a clue about implementing agentic AI. Companies might be blowing through resources to get relatively little (or no) return because of these kinds of fallibilities. 

It also speaks to the issue of time, something the AI agents have no concept of, but that is incredibly important for humans. AI agents cannot experience time because they have no capacity for experiences at all. This leads to funny interactions, like spamming Julia, the intern, in a blitz of ‘did you send it yet?’ pile on emails, sent seconds apart from each other, to the opposite extreme of not accomplishing anything of consequence over weeks. Agents exist in a ‘temporal vacuum’, as Maty puts it, and that is a serious limitation.

There is an enormous amount of time wasted by the humans too, performing human like interactions with agents. It’s the same time suck we have with our human co-workers. However, it doesn’t have the same social payoff as talking with an actual human. It’s exhausting and exasperating in new ways as both Ratliff and Julia learn. 

Reporting to AI

Julia is able to game the agents - telling them that she is not an intern on a temporary contract but the fulltime Social Media Manager - obfuscating the terms of her employment. She insists that she has sent Megan work when it appears she has not sent it at all, according to Ratliff who double checks. She is able to collect a paycheque for a month without having done any actual work because she exploits the AI agents’ lack of memory and understanding combined with the human empathetic traits that Ratliff has assigned them via their backstory. Julia is in the dark about Ratliff’s existence - another level of deception committed in the name of protecting the façade of an AI run organization. 

Julia’s experience raises another ethical issue - will some humans wind up working for AI agents? The tacit assumption when we talk about AI in the workplace is that humans will remain in charge, that people might even command a team of bots. However, this may not be true for all people. Those with less power might be left in roles where they augment the AI driven process. We have already seen this happen with gig workers who essentially perform their work at the behest of an algorithmic system which can assign them a job or take away opportunities. This new deployment merely takes things to another level.

Mesh and The Matrix

“Many leaders still view AI through the narrow lens of automating existing, static, linear processes. This strategy frequently manifests as the creation of “persona-based” agents - tools designed to mimic or replace a specific human job - which misses the true value of agentic AI because it digitizes organizational silos rather than removes them.”  - HBR, A Blueprint for Enterprise Wide Agentic AI Transformation

What would it look like to organize by tasks and outcomes? 

The quoted blueprint post is SPONSORED content from Google  - so take it as the sales pitch that it is, rather than a piece of HBR reporting or research. That said, I think its still useful to illustrate the idea of where this might go. The blueprint goes on to describe a mesh, the architectural layer to coordinate agents across the enterprise. When it comes to organizing AI, mesh might be the new org chart.

Humans organize by role, then task. But, the same task can easily exist across different roles. To take an obvious example, writing reports crosses pretty much every department from marketing to HR to legal to accounting. That task can effectively be accomplished by a single type of agent which is good at writing reports. It could be ‘enhanced’ for context - say legal vs marketing - but the underlying task is the same.

Another fictional example: In The Matrix, Agent Smith, the primary antagonist, is an AI program designed to enforce law and order. When one Agent Smith is annihilated, another replicates seamlessly to take its place. All Agent Smiths have the same capabilities and shared knowledge. They operate more like a hive mind than individual actors. This instantaneous shared knowledge is a huge benefit because it effectively eliminates coordination costs. We might consider Agent Smith to be some kind of artificial general intelligence and while this is all fictional, I think we can take a cue from this approach, minus the anthropomorphizing. 

Agentic AI can be deployed across the siloed domains we have created in the name of coordinating work between humans. This is the enterprise model we should be trying to implement rather than using Agentic AI to speed up or supplement processes siloed within roles. Again - I’m not saying this is a normatively good idea - I’m commenting purely from a business strategy perspective. 

VIPs demand humans

There is another interesting exchange towards the end of the Shell Game that deserves some attention. The CEO agent, Kyle, is meeting with various VCs to fundraise for the company. These high powered people have precious little time for AI. They hang up on Kyle or demand to talk to the ‘real’ founder. These are people hoping to make untold billions by investing in agentic AI, but they themselves will not tolerate it! This holds a clue for businesses that might seek to serve VIP audiences. 

Agentic AI might be fine for the masses, but not for frontline exchanges with high powered people who will have the means to hire real humans. Again, this isn’t so different than how things are done now. You might have a Roomba, but someone with greater means might have Molly Maid and a truly wealthy person has a team of dedicated staff. The irony is that these folks are betting that the rest of us will be just fine with simulated, counterfeit people. Perhaps, in a world of AI agents in workplaces, many of us will need to become like Julia, identifying the weaknesses in the system and exerting human agency to our advantage where we can. Maybe that is the ultimate AI literacy skill.

By Katrina Ingram, CEO, Ethically Aligned AI

Ethically Aligned AI is a social enterprise aimed at helping organizations make better choices about designing and deploying technology. Find out more at ethicallyalignedai.com     

© 2026 Ethically Aligned AI Inc. All right reserved.