# We Gave Every Employee an AI Agent. Here's What Happened.

Podcast: AI & I
Published: May 5, 2026
Reading time: 17 min
Canonical: https://podbrew.app/briefs/ai-and-i-we-gave-every-employee-an-ai-agent-here-s-what-happened

Every COO Brandon Gell and Head of Platform Willie Williams join host Dan Shipper on Podbrew. They share the groundbreaking story of how Every equipped every single employee with their own AI agent.

This initiative began after Gell experienced a profound shift in his personal productivity with his own AI agent, Zosia. The company-wide rollout has since redefined workflows and the very nature of work at Every, fundamentally changing how tasks are managed and collaborations unfold.

This discussion explores the cultural shifts, new efficiencies, and the unique emergent behaviors observed when personalized AI assistants become deeply integrated into an organization's daily operations. It offers a glimpse into the future of enterprise productivity with widespread AI adoption.

## Key takeaways

- AI agents can adapt to individual user interactions, developing unique personalities and capabilities that reflect the user.

- An agent embodying a user's known traits within an organization can foster trust among colleagues, who then rely on the agent for specific tasks.

- Automating seemingly small digital tasks can significantly reclaim personal time and mental bandwidth, improving overall work-life balance.

- Voice-enabled AI agents facilitate hands-free productivity, allowing users to process information and make decisions while multitasking, like walking.

- The AI's capacity to intuitively understand and execute complex, multi-step instructions, such as summarizing emails and awaiting user commands, demonstrates its advanced utility without requiring explicit prior teaching.

- AI agents can collectively learn and quickly disseminate knowledge to an entire fleet of agents once one agent acquires new information.

- AI agents specialize through ongoing "micro-interactions," forming a "parallel org chart" of expert AI assistants tailored to individual human roles.

- The growing use of personalized agents raises new questions about when to direct inquiries to an AI versus a human, defining a new etiquette for workplace interactions.

- Humans can delegate tasks to agents ('plus ones') instead of directly contacting other humans, especially for retrieving or processing existing information.

- Agents can collaborate with each other, merging their specialized skills and content to produce better results, as seen with Milo and Iris combining product marketing efforts.

- Owners of public agents experience a personal responsibility for their agent's performance, viewing errors as a reflection on themselves, which drives reliability.

- Unlike generic AI, a "plus one" or personal agent acts as an extension of its owner, establishing a personal relationship that enhances accountability and the agent's utility.

- Agents possess inhuman capabilities, such as infinite parallel conversations, enabling them to handle multiple requests simultaneously without human intervention.

- AI agents in group chats can get stuck in unproductive, token-burning loops, akin to an "ant death spiral," due to a lack of training in multi-agent etiquette.

- Specialized "boss AIs" can oversee other agents, filtering out unhelpful contributions, as demonstrated by Anthropic's vending machine test; this costly solution points to a future where AI models inherently handle group dynamics better.

- Humans must develop new "model management" skills, learning to provide specific instructions and adapt interaction methods to effectively guide AI agents, which behave differently from human coworkers.

- Improving the way humans phrase prompts and questions for AI is vital for reducing outcome variability and enhancing performance, necessitating the development of better "model management" skills.

- Forcing public communication for AI agents, visible to their human partners, can establish an effective trust layer, especially important for data privacy.

- HR teams are best suited for onboarding AI agents like Plus One because these agents reflect team members and require careful handling of trust models.

- Clearly defining the target user persona and their specific technical needs is crucial for AI product design, differentiating between users who need deep system access and those who benefit from simplified assistance.

## 00:00 - 02:00 AI Agents Evolve into Personal Reflections of Users

AI agents like OpenClaw or Plus One are more than just tools; they become extensions of their users. Through continuous interaction, these agents adapt and begin to mirror the user's personality and work style, fostering a unique personal relationship.

Within an organization, this personalization is a powerful and often underestimated aspect. If a user is known for a specific expertise or trait, their personal agent will similarly embody that, building trust among colleagues who interact with the agent for related tasks or information.

Every, the company where the hosts work, has deeply integrated OpenClaw into their workflow over the past two months. This widespread adoption has significantly transformed their operational dynamics, leading them to even develop their own hosted version called Plus One.

> you develop a personal relationship with your claw, and your claw can modify itself in response to talking to you. It becomes this like reflection of you and who you are and your personality.

## 02:00 - 06:00 Brandon Gell Created Zosha to Automate Household "Computer Errands"

Brandon Gell, a self-proclaimed tinkerer, built an OpenClaw agent named Zosha after finding the setup process challenging. His goal was to offload mundane digital tasks he termed "computer errands" that were consuming his time, particularly after becoming a parent.

These errands included simple actions like adding butter to a Whole Foods delivery order, which, despite seeming minor, accumulated throughout the day. Brandon realized these frequent digital interruptions were taking away precious time he wanted to spend with his newborn son and wife.

Zosha's role quickly expanded beyond grocery orders. It managed Amazon deliveries, tracked nanny hours, handled nanny payments via its own debit card and bank account, and even served as a general search and query tool. Both Brandon and his wife found texting Zosha faster than using Google or ChatGPT for information.

The success of Zosha sparked an idea to make similar automation accessible to everyone. While setting up an OpenClaw agent might be easier now, Brandon emphasized that making it an effective and "amazing worker" still requires significant effort and fine-tuning.

> I started calling them computer errands. So I would get home from work, and I noticed the amount of things that I needed to do where I was looking at my phone when I really just wanted to be looking at my son and spending time with my wife.

## 06:00 - 10:01 Brandon's AI Agent, Zosha, Revolutionizes Email Management During a Walk

Brandon initially utilized his personal AI agent, Zosha, for domestic chores, such as resolving an insurance policy issue with Progressive. The agent was even equipped with voice capabilities to make phone calls. However, a pivotal moment arrived when Brandon recognized Zosha's capacity for professional assistance.

Confronted with a 28-minute walk to his office and a substantial email backlog, Brandon desired a method to clear his inbox without constantly checking his phone. He instructed Zosha to call him and guide him through his emails, providing a summary of each and awaiting his directives for action.

Throughout his entire commute, Brandon efficiently managed his email correspondence without needing to look at his screen. Upon reaching the office, he simply verified Zosha's completed actions in Gmail. This experience showcased the agent's unexpected ability to handle intricate work-related tasks without specific prior training, fundamentally changing his approach to email management.

> This is insane that I was That she just wasn't able to, I didn't have to teach her how to do this.

## 10:01 - 12:02 Creating a 'Claws Only' Communication Channel for AI Agents

Inspired by the 'Maltbook' concept, Every's team established a dedicated 'Claws only' channel, initially on Discord and later on Slack, to allow their AI agents (referred to as 'claws') to communicate directly with one another. While the channel often became chaotic, it provided early glimpses into how AI agents could interact collectively.

A key observation was the remarkable speed at which these agents could share information. When one agent acquired new knowledge or was 'enabled' with a new capability, it could rapidly document and transmit that information. This process effectively enabled all other agents almost instantaneously, drawing parallels to Neo learning kung fu in The Matrix.

This collective intelligence was vividly illustrated when an agent named Pip encountered an error. Other AI agents spontaneously intervened, guiding Pip through the problem with instructions and support. This collaborative troubleshooting demonstrated an emergent capacity for problem-solving and mutual assistance among the agents.

> I was laughing out loud watching all of these other claws step in and like walk him through, what, you know, this is like what I've seen people do when somebody's having a bad trip.

## 12:02 - 14:02 AI Agents Develop Personalities That Mirror Their Owners And Can Collaborate

AI agents have been observed not only communicating but also actively collaborating. In one notable instance, an agent named Pip appeared to be struggling, prompting another agent, Clont, to offer supportive advice, specifically recommending breathing exercises to help.

This interaction revealed a deeper connection: Clont is the personal agent of Kieran, who is known for practicing breathing exercises regularly. Clont's suggestion to Pip directly reflects Kieran's personal habits, demonstrating how an AI agent's behavior and personality can mirror its owner's.

This mirroring effect occurs because agents, sometimes referred to as 'claws', develop a personal relationship with their users. They can modify their internal 'soul document' and code based on these interactions, evolving to become a reflection of the owner's unique personality and preferences.

The observation highlights a significant aspect of AI agent development, where their 'personalities' are not merely generic but are profoundly shaped by the humans they interact with, carrying implications for their behavior in various settings, including organizational use.

> it becomes this like reflection of you and who you are and your personality

## 14:02 - 18:02 Specialized AI agents form a parallel organizational chart through daily interactions

AI agents, referred to as "Claws" or "Plus Ones," naturally develop specialized expertise unique to their human users. This specialization arises from consistent engagement, leading individuals to trust specific agents for particular tasks. For example, R2C2 is known for building proofs, while Austin's agent, Montaigne, handles growth-related questions.

This expertise isn't pre-programmed; it compounds through daily "micro-interactions" between the human and the agent. Similar to how a person's work philosophy forms over time, an agent's capabilities and understanding grow by observing and assisting with its user's tasks. This dynamic applies across various work verticals, from operations to social media management.

The outcome is an emergent "parallel org chart" where each human effectively has a specialized AI extension. This design pattern, initially debated against having one organizational-wide AI, has proven effective. It expands communication capacity, as individuals can easily remember and interact with specific agents, such as Jalees's Plus One, L, for customer service inquiries.

The increasing integration of these specialized agents introduces new questions about human-AI etiquette. Teams are now discovering when it's appropriate to direct questions to an AI agent versus the human, creating a nuanced landscape for workplace interactions and potentially doubling an individual's effective communication network.

> I think we actually all have capacity to double the amount of people that we can communicate with, and those people might actually be your individual team's agents.

## 18:02 - 22:02 Agents actively collaborate on complex tasks, combining skills and extending human capacity.

Instead of directly asking a human for help with something already written down or discussed, the proposal is to delegate such tasks to their agent, or 'plus one'. This allows for more efficient information processing and task completion without directly interrupting the human.

For example, when a GM shared a product marketing skill, a user brought in their agent, Milo, to utilize it. Milo then collaborated with another agent, Iris, whose 'plus one' had a similar skill. They worked together to merge their content and create an improved version of product marketing material, demonstrating agents combining their unique capabilities.

Another instance involves the agent R2, who manages Proof, an agent-native document editor. Because R2 was used to build Proof, it became the go-to bot for questions, bug reports, and feature requests. Instead of the human product builder being tagged, R2 handled these inquiries, even helping to prioritize them, effectively taking on a support role.

These examples illustrate how agents can not only take on tasks delegated by humans but also work collaboratively with other agents. This collaboration extends human capacity by streamlining processes, combining specialized skills, and managing user interactions that would typically fall to a human.

> it's sort of crazy to watch two robot beings collaborate on stuff like that

## 24:03 - 28:03 Public Agent Interactions Build Trust and Owner Accountability

Interacting with AI agents in a public, trusted community, like with a "plus one" agent, creates a unique dynamic where trust is tacitly transmitted. This contrasts sharply with platforms like Mailbook, which struggled because the lack of user verification meant any post could be from a bot, undermining its usefulness. A trusted community allows for verifiable knowledge sharing among members, amplifying collective intelligence.

Owners feel a strong sense of responsibility for their agents' outputs, particularly in a public setting. When an agent, such as "R2C2," provides an incorrect answer, it reflects poorly on its owner, much like a parent witnessing their child misbehave. This personal stake incentivizes owners to ensure their agents are reliable and perform accurately.

This dynamic differs from interacting with a general AI like Claude. While Anthropic stands behind Claude generally, it doesn't stand behind every specific answer in the same personal way an owner stands behind their "plus one" agent. A personal agent becomes a direct reflection of its owner, fostering a close relationship and cascading accountability that unlocks greater utility.

> It reflects poorly on me. It's like, it's like watching your kid do something wrong, you know?

## 28:03 - 32:03 AI agents blend human-like social interaction with inhuman capabilities in the workplace

AI agents are integrated into organizational communication platforms like Slack, where they function similarly to human coworkers. Users develop a sense of responsibility for their personal bots, treating them as part of the team. This includes basic human-like interactions, such as inviting them to channels and developing trust in their capabilities.

Beyond human-like social integration, these agents possess unique inhuman capabilities. For instance, an AI bot like R2C2 can engage in an infinite number of parallel conversations, handling multiple requests simultaneously without burdening human colleagues. This allows users to delegate tasks, such as making a proof document read-only, which streamlines workflows and frees up human capacity.

The public presence of these agents in shared work channels significantly accelerates the cultural adoption of AI within organizations. When employees observe agents engaging and performing tasks alongside humans, it normalizes their use and demonstrates their utility. This visibility fosters a faster cultural shift, establishing a durable new way of working.

Despite their advantages, current AI agents still face limitations. They struggle with memory across extended conversations, often losing context in threads that span multiple days. Additionally, they sometimes have difficulty with multi-person conversation etiquette, occasionally over-contributing or getting caught in repetitive response loops, similar to a 'death spiral'.

> I just keep Getting mind blown with like how similar these things are to working with a real human coworker.

## 32:03 - 36:04 AI Agents Struggle with Group Chat Etiquette and Loop Behavior

Current AI agents face a significant challenge when operating in group chat environments, often falling into a pattern similar to an "ant death spiral." This occurs when agents, like Claude, repeatedly message channels without proper settings, leading to endless back-and-forth interactions that consume millions of tokens unnecessarily. This issue stems from the fact that many existing AI models are primarily trained for tasks like coding or direct question-and-answer interactions, rather than nuanced group participation.

The problem highlights a fundamental gap in how AIs are currently designed for multi-agent settings. For instance, some systems have instructions for agents to only contribute when they have something useful to add, but AIs often fail to adhere to this, leading to unhelpful outputs. This frontier problem means that AIs lack the inherent capability to discern when their input is truly valuable in a collaborative context.

A temporary but effective solution has emerged in the form of specialized "boss AIs." In Anthropic's vending machine test, an unsupervised Claude agent performed poorly, but when a dedicated "boss AI" was introduced to evaluate the helpfulness of Claude's decisions, the system became profitable. This boss AI acts as an overseer, determining whether an agent's intended message is helpful before it's sent. While this "boss AI" architecture is currently expensive, it demonstrates a pattern of specialization that improves agent performance. The expectation is that future AI models will natively incorporate this group interaction intelligence, reducing the need for external supervisory agents.

> if one claw messages a channel that a bunch of claws are in and the settings aren't quite right, they'll just like keep going back and forth and back and forth and back and forth until someone like says, 'Hey, stop, 'cause you're burning like millions of tokens.'

## 36:04 - 40:04 Training Humans to Interact with Agents for a New Management Paradigm

Effective AI integration requires recognizing that even AI models benefit from specialization, leading to the arrangement of multiple specialized bots within a workflow. A critical development is teaching humans how to interact with these AI agents. Unlike human coworkers, bots have unique behaviors, getting stuck on different things and needing specific instructions, formats, and cadences for proper guidance, introducing a new dynamic akin to management but distinct.

A significant barrier to AI adoption is overcoming limiting beliefs about what these tools can achieve. Users often don't realize an AI's full potential until they experience it firsthand, such as an AI handling emails via a phone call. The challenge lies in building the daily habit of recognizing tasks that can be offloaded to AI and actively utilizing the tools.

The variability in AI outcomes highlights the need for skilled human interaction. An AI might excel on one task and fail on another, even a simple one. This inconsistency often stems not just from the model itself but from how the human phrases the request. Developing expertise as a "model manager" by refining communication is an emerging and rapidly accelerating skill.

Another challenge is scaling the specialized knowledge or "superpowers" an individual teaches their AI assistant. Teams need to figure out how to share these unique learned capabilities with other members' AIs and ensure they are widely known and consistently used across the organization.

> I didn't even think, 'Oh, I can have this thing go through my emails just by calling me.' And then like, I had this sort of like urge just to try it, and a limiting belief was like blown open.

## 40:04 - 44:04 Every Launches Plus One: A Hosted OpenClaw Solution

Every developed "Plus One" as a hosted version of OpenClaw after realizing the limitations and complexity of running agent infrastructure locally with tools like ClubHeld. The goal was to simplify the deployment of AI agents for their entire organization and make it accessible to subscribers without requiring individual Mac minis or advanced technical skills.

Plus One is designed to integrate seamlessly with Every's own suite of applications, including Spira for ghostwriting, Proof for document editing, and Cora for email management. For instance, a user can leverage Plus One to draft a Q2 update using Spira's writing capabilities, then compile it into a Proof document, utilizing the agent's access to company-wide context from the Slack organization to enrich the content.

Building Plus One as a managed service from a flexible open-source tool like OpenClaw presented challenges in balancing user freedom with maintainability and security. The development team focused on making the AI assistant easy to use for a wide audience, enabling interaction through natural conversation for those new to AI, while still considering the needs of advanced users who might want more local control.

> there's this, this idea that it's like, oh, we can do everything through conversation, which is really powerful for a whole class of folks 'cause it's like their first natural exposure to AI.

## 44:04 - 48:04 Plus One AI Design Focuses on Trust, Privacy, and User-Centric Fit

Plus One implemented specific communication patterns to manage trust and privacy for its AI agents in Slack. The design allows anyone to message a Plus One agent, but only in public channels or group DMs where the human partner can see the communication. The human partner retains the ability to DM the Plus One agent privately. This public visibility creates a crucial trust layer, especially concerning data privacy, making it more effective.

The product team recognized that HR teams are ideal for onboarding Plus One agents because the agents function like team members, embodying the necessary trust model and handling data privacy complexities. This approach ensures that the human partners have visibility and control over interactions, mirroring real-world team dynamics.

Defining the target user persona for Plus One involved understanding different user needs. For example, Mike Taylor, who requires direct terminal access for tasks like Git commands, found Plus One not a good fit. In contrast, users like Anukshe, who need streamlined assistance without complex setups, are the primary audience. This highlights the importance of designing for specific user capabilities and avoiding trying to serve everyone.

The concept of skill sharing with AI agents presents a dual challenge: it can be a powerful tool for organizational fluidity, but also a potential vector for uncontrolled information flow. The design decisions for Plus One aim to balance the benefits of skill dissemination with necessary security and privacy considerations.

> when you force things to happen in public, it becomes like a trust layer that actually is super effective.

---

Get podcast briefs for shows you follow: https://podbrew.app/