Aaron Levie says agents will soon outnumber humans 100x to 1

Agents Arrive Before the Internet Is Ready

The hosts open by arguing that AI agents are no longer a speculative concept but an architectural shift that software companies need to treat as immediate. They center the discussion on a post by Aaron Levie, who says the industry has reached an inflection point: “agents are becoming the primary users of software,” even though the web’s pricing models, standards, and enterprise systems were built for humans. The core prediction is stark. Very soon, there may be “100x” or even “1,000x more agents than people” interacting with software. If that happens, every assumption behind modern product design changes.

Nick summarizes the thesis with a new maxim: product builders once followed Paul Graham’s advice to “make something people want,” but now they may need to “make something agents want.” That means software must become API-first, structured, and machine-readable. Clean codebases, updated documentation, and reliable interfaces cease to be nice-to-haves; they become prerequisites for agent adoption. Human-first growth tactics also weaken. An agent will not attend a webinar, click a marketing email, or browse a landing page for pleasure. It will choose the tool that works best and move on.

Matt agrees with the premise and draws a sharp contrast between business software and consumer software. In B2B, many products already expose APIs because they want developers and partner platforms building on top of them. “APIs are the language of agents,” he says, which gives enterprise software a head start. Consumer software is the opposite. It is “completely broken” for agents because publishers, e-commerce sites, and consumer services have spent 20 years trying to block bots.

That tension produces one of the show’s clearest insights: the internet is being approached from two directions at once. On one side, startups and model providers are improving browser control and computer-use agents so they can navigate human interfaces despite friction. On the other, agent companies are lobbying websites and platforms to expose direct interfaces for machine interaction. The hosts expect those two approaches to “meet in the middle,” but the winner will likely be whichever consumer platform first offers trusted agent authentication and robust API access.

The result is a reframing of the next software era. The hosts are not talking about better chatbots. They are talking about a rebuilt digital economy in which products compete not only for human attention, but for machine preference.

The Consumer Web’s Incentive Problem

After establishing that the internet must be rebuilt for agents, the hosts focus on the hardest part of that transition: the consumer web. Matt argues that the biggest obstacle is not technical capability alone, but incentives. In business software, API access often aligns with vendor goals. In consumer software, the dominant revenue model still depends on human eyeballs. If agents browse on behalf of people, many publishers and platforms lose ad impressions, page views, and direct merchandising opportunities.

Amazon becomes the clearest example. Matt calls it “crazy” that a personal agent still cannot simply buy products through Amazon without navigating the visual website like a human. The same problem exists with credit card portals and paid media subscriptions such as The Information or The Wall Street Journal. Users may already have access rights, but their agents still cannot reliably exercise those rights without browser gymnastics. That makes many real-world consumer tasks frustrating, slow, and brittle.

Nick believes Amazon and similar platforms will eventually adapt because the innovation wave is too large to resist forever. Matt pushes back with a more nuanced economic argument. Amazon makes “a ton of money from advertising,” and publishers depend heavily on ads as well. An agent that completes a purchase or retrieves an article without displaying banners removes the monetizable moment. In that world, the platform may gain transaction efficiency but lose attention revenue. That creates a natural reason to delay adoption until competitive pressure becomes overwhelming.

The hosts see this tension as one of the defining strategic questions of the next few years. Consumer companies may not move first because they want to; they may move only when enough personal agents exist that refusal begins to cost them sales. Until then, the transition is likely to be uneven. Some agent companies will keep improving browser control, while others try to strike business-development deals with publishers and e-commerce platforms to expose direct interfaces and standardized purchasing paths.

The segment also reveals why enterprise use cases still dominate agent adoption. Matt says most of his use with OpenClaw has been business-related because business software is simply more legible to machines. Consumer interfaces remain hostile, anti-bot, and rarely optimized for authentication by proxy. That means the first big agent productivity wins are happening inside work, even though the long-term opportunity on the consumer side could be larger.

The takeaway is not just that consumer software lags behind. It is that the consumer internet’s business model was optimized for human attention, while the agentic internet will be optimized for delegated outcomes. Those are not the same thing, and the collision between them is only beginning.

Documentation Becomes Survival Infrastructure

The hosts then turn to a less glamorous but more foundational issue: documentation. Nick frames it as a “boring problem” that may nonetheless become a bottleneck for deploying agents at scale inside enterprises. Historically, when processes were undocumented, employees could compensate by asking colleagues, tapping someone on the shoulder, or learning through informal tribal knowledge. Agents cannot do that naturally. If workflows, standard operating procedures, and decision logic are not written down, the risk is that agents “flail” without the right guardrails.

Matt’s answer is blunt. “Any company that is not documenting their processes is inefficient and likely bound for failure.” His argument is broader than AI. If a company’s expertise lives mainly in people’s heads, it becomes fragile. When those employees leave, the organization loses memory, continuity, and speed. Agents merely expose that weakness more quickly. Documentation is not just helpful for automation; it is the storage layer for institutional knowledge.

This point links directly back to the earlier thesis about rebuilding software for agents. Structured data and APIs matter, but so does narrative knowledge: how teams really work, how exceptions are handled, what context matters, and which steps are critical. Matt describes this accumulated knowledge as the “memory” and expertise built up within an organization over time. To make agents effective, companies need to “shortcut the onboarding process for a new agent” by embedding that organizational memory into systems the agent can access.

Nick responds with a candid admission that his own team likely has not documented enough. That moment gives the segment practical force. The problem is not abstract and it is not confined to legacy enterprises. Even AI-native teams often operate with informal processes, making them vulnerable to inconsistency and hard to scale.

The conversation suggests that documentation will evolve from static wiki pages into active machine context. It will need to be well-organized, current, and tied to real workflows rather than aspirational process charts. The companies that treat documentation as a living operational asset will be in a better position to train internal agents, preserve institutional memory, and improve execution quality.

The hosts imply a simple but powerful progression. First, firms digitized communication. Then they digitized workflows. Now they must digitize judgment and context. Documentation sits at the center of that shift. In a world where agents are expected to perform meaningful work, poorly documented organizations are not just inefficient. They are unreadable.

OpenClaw Sparks the Personal Computer Agent Race

The second news topic centers on Perplexity’s launch of “Computer,” a cloud-based agent that orchestrates “20 different AI models” to execute complex, multi-step workflows autonomously. Shortly after, the company announced “Personal Computer,” a persistent system running on a Mac Mini. The hosts read this as a direct response to the rise of OpenClaw, an open-source computer-use agent built by Austrian developer Peter Steinberger.

Nick summarizes OpenClaw’s improbable impact. The project went “incredibly, incredibly viral,” was later “acquired by OpenAI,” reached “200,000 GitHub stars,” and reportedly helped sell out Mac Minis as hobbyists and power users rushed to run their own local AI assistants. In his view, OpenClaw forced the rest of the market to react. Perplexity, Tencent with “Co-work,” and Cursor with its automation efforts all appear to be converging on some version of the same idea: a persistent personal AI that can act across software and operating systems.

Matt sees this as a positive development. He says OpenClaw inspired both hobbyists and major companies, and he is excited to test Perplexity’s version. But the two hosts differ on how mainstream these systems are becoming. Nick thinks products like Perplexity Personal Computer may have a shot at “the masses,” precisely because they remove OpenClaw’s installation burden and security rough edges. Matt is more skeptical. He agrees the audience is broader than OpenClaw’s, but he still does not imagine his parents adopting it soon. The interface remains “a little too intimidating,” too dependent on prompt craft, experimentation, and technical confidence.

That disagreement leads to an important market distinction. These agents may become highly valuable first inside work settings, where users already have clear tasks, higher motivation, and software-heavy routines. Everyday consumers, by contrast, still need obvious use cases and simpler onboarding. Matt thinks simple question-answer products like ChatGPT, Claude, and Perplexity are still better aligned with mainstream habits than persistent computer agents.

The hosts also joke that every AI company is effectively saying, “I guess we’re building OpenClaw now.” The line captures the broader point: OpenClaw may not dominate the category forever, but it has already shaped the roadmap of the industry. Even tools like Claude Code have, in their view, shipped updates in the last two weeks that feel increasingly “openclawesque.”

The takeaway is that the personal assistant race is no longer theoretical. OpenClaw proved the appetite, Perplexity is packaging it for broader users, and nearly every major AI company now appears to be pursuing a version of the same future.

Slack Wants to Become the Operating System for Agents at Work

The first guest, Rob Seaman, EVP and GM of Slack, argues that Slack is evolving far beyond team chat. He says the company now positions Slack as “an operating system for the agentic enterprise.” The phrase is deliberate. Traditional operating systems hide hardware complexity and provide interface rules for applications. Slack wants to play a similar role for enterprise agents: a place where they can be distributed, coordinated, and interacted with as naturally as coworkers.

Seaman invokes an old internal line from Slack co-founder Stewart Butterfield, who used to say Slack was “a search and AI tool disguised as a messaging tool.” In his telling, that idea has finally become real. Slack’s advantage is context. The platform is made of channels, and those channels mirror what a company cares about: major projects, priorities, launches, teams, deadlines. At Salesforce, he says, there are channels tied to all the important goals for fiscal year 2027. At an individual level, people also have channels around their own next “30, 60, 90 days” of work. That means Slack contains both organizational and personal context in a shared graph.

He argues the company was “built for this moment” because Slack’s architecture resembles a social network more than traditional enterprise software. The challenge is not simply storing all messages from companies that may have been using Slack for “10 plus years.” It is selecting the context that matters in the moment and feeding that into the model. Even as context windows grow, Slack sees its job as identifying the most relevant company and user history to surface.

The hosts zero in on why this matters strategically. If agents can work effectively in Slack, more actual work can move into Slack rather than merely conversations about work. Seaman agrees and says that pattern is already visible. In Slack’s earlier era, people discussed work in channels, then jumped into other apps to approve expenses, process offers, or complete tasks. Apps brought lightweight actions into Slack, but “thicker work” still happened elsewhere. Agents change that. Now, if a user is in a customer feedback channel, they may be able to trigger Cursor or Linear directly inside Slack and implement a change without leaving the conversation.

That is the core strategic shift: Slack stops being a wrapper around communication and becomes a control plane for execution. If successful, the app becomes the place where people, apps, and agents all coordinate work in one environment. In that world, the messaging layer does not disappear. It becomes the natural interface for enterprise AI.

Slackbot, Skills, and the Fight Against Channel Bloat

One of the most concrete product discussions of the show focuses on Slackbot and the problem of “channel bloat.” Nick admits that one of his main frustrations with Slack is the proliferation of channels: one for every initiative, campaign, team, and department until the sidebar becomes unmanageable. Seaman responds that the problem is real and “not your fault.” He says Slack has wrestled with this from the beginning, but AI now makes it possible to solve more effectively through summarization, sorting, and semantic retrieval.

Slackbot is central to that vision. Seaman describes it not as a public, autonomous participant in channels, but as “your personal agent.” Users no longer need to know exact channel names or remember where a conversation happened. They can ask natural-language questions like, “Where was I talking to Nick and Matthew about the evolution of Slack like two weeks ago?” and Slackbot can retrieve the right thread semantically. Seaman says he does not even look at his sidebar anymore because Slackbot finds what he needs.

Privacy and permissioning are handled through user-level access. Slackbot only accesses what the user can access. Seaman gives an example from internal all-hands meetings. He asked Slackbot to find the presentation deck and figure out how to pronounce the names of new hires. The agent located a private Slack channel he belonged to, queried his connected Google Drive, found the deck, extracted the relevant names, and generated pronunciation guidance. Crucially, it did all of this using his permissions, not some global company privilege. Slackbot can search private channels, public channels, and connected systems, but only within the boundaries of the requesting user’s auth.

The interview also touches on the product’s viral adoption. Inside Salesforce, Slackbot reportedly went from zero to “50,000 weekly active users within the first few weeks,” and Seaman says the company “didn’t even tell anybody about it.” Usage spread through what he calls “hyper local social proof,” with coworkers sharing effective prompts inside their own functions. A salesperson learning a useful prompt from another salesperson, he argues, is much more powerful than generic best-practice sharing across companies.

Slack is also developing reusable prompt logic under the label “skills.” These skills can be created by individuals, shared with colleagues, or standardized by IT and business teams for repeatable workflows like account planning or agency creative reviews. Seaman says the company has already deployed this internally and will discuss it publicly at an event on March 31.

The broader point is that Slack is trying to turn retrieval, prompting, and workflow execution into a native part of work. The company is not simply layering AI on top of chat. It is trying to reduce navigational friction, codify expertise, and let users summon organizational context with natural language.

Slack’s Internal Use Cases and the Economics of Enterprise AI

The most revealing part of Seaman’s appearance comes when the hosts ask for actual use cases. He responds with examples that show how deeply Slackbot is being used inside Slack itself. Every product feature has its own public internal feature channel, he says, and the Slackbot feature channel now has “3,000 people” actively watching and contributing ideas. He routinely asks Slackbot to summarize the last 30 days of feedback, split it into positive and negative themes, remove employee names while keeping their roles, and then transform that information into different outputs.

One example becomes a mini case study in modern knowledge work. Seaman asks Slackbot to turn positive feedback into “a series of seven tweets” for Slack’s March 31 event, then generate one version in the voice of “Slack HQ” and another in the style of Marc Benioff. He can also take the negative feedback and ask for a product brief for engineering or a sprint plan for the next two cycles. In other words, one corpus of internal conversation becomes raw material for marketing, product management, and engineering prioritization without manual synthesis.

Another example comes from pricing. A team preparing a product price change gave Slackbot the deck, then asked it to connect that proposal to Salesforce customer records and Slack conversations about each account. The goal was to predict how specific customers might respond and recommend what account executives should do. Seaman says the system scanned “thousands of customers” to produce a targeted action plan.

These examples support the hosts’ larger thesis that AI’s biggest immediate value may be compressing the distance between unstructured communication and executable work. Feedback no longer sits idle in a channel. It becomes strategy, messaging, and task planning at machine speed.

The hosts then turn to the cost question. As use cases become more sophisticated, they require more tokens and better models. How does Slack think about pricing when demand could become enormous? Seaman says Slack has chosen a relatively simple model so far: Slackbot is baked into plans with usage limits tied to tiers rather than sold as purely metered consumption. Some especially expensive features may eventually be reserved for higher plans or charged by usage, but the company wants the experience to remain “dead simple” for enterprise buyers.

That pricing philosophy reflects Slack’s broader strategy. Rather than making AI feel like an add-on toll booth, the company wants it to feel like a native part of the product. The bet is that if agents become central to work, Slack’s value expands far beyond messaging. In that future, charging only by token usage would undersell the platform’s role as enterprise infrastructure.

Giga Bets on Voice Agents and the Automation of Support Work

The second guest, Giga co-founder and CEO Verun Vermati, explains why his company abandoned its original business of fine-tuning LLMs for enterprises and went all in on voice AI agents. The shift, he says, was driven by market reality. Model providers kept dropping prices, increasing throughput, and improving base performance, which weakened the need for custom fine-tuning as a standalone offering. The two areas where customer demand remained strongest were support and coding, and Giga chose support.

He describes Giga as a platform for customer support agents that work across “voice, chat, email” and other channels. Customers include DoorDash, where Giga helps power support for delivery workers. The company raised “$61 million” at the end of the previous year. Vermati says Giga supports both frontier models and custom fine-tunes, depending on customer requirements. In regulated industries, some buyers still prefer non-frontier or more controlled setups. In other cases, Giga fine-tunes models for voice naturalness, such as removing the tendency to answer in formal numbered bullet points rather than conversational speech.

One of the most striking claims in the interview is around resolution rates. Nick says he has heard Giga can eventually reach “98% resolution,” though Vermati clarifies that implementations often start far lower, around “15 to 20%.” The reason is not that the models are weak. It is that most companies have poor support documentation. Policies are scattered, contradictory, or trapped in the heads of operations staff. Giga begins by ingesting whatever exists, studying transcripts from the best human agents, and proposing policy changes. Then, whenever calls are escalated, it analyzes what the human did that the AI could not do and suggests further updates.

That creates an iterative learning loop. Resolution improves not just because models get better, but because companies are forced to formalize their support logic. In some cases, automation also requires new APIs or new capabilities because humans have access to systems the AI cannot yet reach. Vermati says Giga has even launched browser agents to close some of those gaps faster.

The workforce implications are acknowledged directly. He says AI support adoption is often tied to opex reduction, but he also argues that automation creates new roles around policy management, system improvement, and oversight. More importantly, he insists AI enables classes of support that humans could not economically provide at all. That claim becomes central in the next section, where he argues that AI should not only replace repetitive support tasks but make entirely new kinds of service viable.

Forward Deployed Engineers, OpenClaw Economics, and New Support Possibilities

Vermati’s most provocative argument concerns forward deployed engineers, or FDEs. He defines them broadly as a lighter, customer-facing version of engineers or solutions engineers who work directly with enterprises to understand a workflow, craft prompts, build policies, and make AI systems function in production. His criticism is not that these roles are useless, but that they are becoming a bottleneck to enterprise AI adoption.

He argues that many AI companies appear to be selling products when, in practice, they are selling a team of humans who sit with the customer for “6 to 9 months” or longer. In one case, he says a Fortune 500 company working with a major model lab spent “more than 12 months” trying to automate support and still could not get “one intent live.” The limiting factor was not demand. Enterprises want support automation. The limiting factor was the manual labor of turning messy transcripts, inconsistent policies, and internal complexity into functioning agent systems.

His conclusion is blunt: “Humans are the biggest bottleneck right now to scaling of enterprise automation.” Giga’s own bet, he says, is to “build an AI forward deployed engineer.” He believes much of the reactive work these teams do, like analyzing transcripts, identifying policy gaps, and iterating prompts, can itself be automated by AI. He distinguishes that from the proactive educational side of enterprise consulting, which may still require human guidance. But the repetitive implementation work, he argues, is exactly the kind of engineering-adjacent task that tools like Claude Code and similar systems will eventually absorb.

The conversation also circles back to OpenClaw. Vermati says the project is “super expensive to run” and may have had a spike of enthusiasm that does not yet translate into sustained mainstream usage. Its security flaws and rough edges limit broader deployment. Matt partly agrees on the friction but insists his own usage is “more than ever,” noting that his token bills are substantial. Both hosts and guest agree on one thing: OpenClaw inspired a whole wave of follow-on products.

Perhaps the most optimistic part of Vermati’s interview comes when discussing customer support itself. He argues AI can create experiences that human teams would never offer at scale. At DoorDash, for example, if a driver says the customer is not answering, the AI can proactively place an outbound call to the customer while the driver remains on the line. That kind of multi-party, real-time coordination would be hard to provide economically with humans. Another customer, a European crypto exchange, uses AI to call users who dropped off during KYC and bring them back into the funnel.

For Vermati, this is the real promise of support agents: not just cost reduction, but service expansion. AI makes it possible to pursue “lost opportunity” that businesses would otherwise leave untouched.

Memory, National Security, and AI Health Companions

The final stretch of the episode combines two of the show’s most forward-looking interviews. First, Letta co-founder and CEO Charles Packer argues that memory, not just model size, will define the next stage of agent development. He frames the problem as “memory” or “learning”: language models are powerful, but unlike humans, they do not truly learn continuously. Knowledge still tends to arrive in new weight releases rather than through dynamic, personalized accumulation.

Packer says the recent breakthrough is that models have become “really, really good at computer use,” which opens a new path: letting agents manage and edit their own memory almost like files on an operating system. He ties this to his ideas around “sleep time compute,” the notion that AI should continue useful cognitive work even when the user is not actively engaging with it. At Letta, that means running reflection agents in the background that analyze activity, extract higher-order insights, and write durable memory rather than just raw logs. His broader claim is that accumulated memory becomes the real moat. Switching between coding agents is easy today because there is little lasting personalization. Once systems build rich, self-curated memory, leaving them becomes much harder.

Packer also raises a national-security warning. As memories become more valuable than model weights, companies and nations may treat them as strategic assets. A future attacker might care less about stealing a model checkpoint than about stealing the “memories” a system has formed about millions of users. Those memories could contain behavioral patterns, psychological insights, and a compressed form of collective knowledge.

The last guest, Microsoft AI’s Dr. Dominic King, brings the conversation into healthcare. King says Microsoft sees “50 million” health-related sessions or queries per day, and health is the number one intent on Copilot for both mobile and voice. About “40%” of those queries are basic health information requests, while roughly “10%” concern symptoms or help interpreting test results. “One in seven” conversations, he adds, are about someone other than the user, such as a family member or friend.

King describes Microsoft’s ambition as building a secure “health companion,” not merely “a doctor in your pocket.” The goal is to combine broad medical knowledge, personal context, connected wearables, and eventually health record data to help users understand symptoms, navigate care, and become more proactive. He stresses that this must be done with layered safeguards, physician oversight, medical policy, and continuous evaluation. The payoff, if done well, is a shift from reactive medicine to more preventive, personalized support.

The episode closes on that optimistic note. Across work, support, memory, and health, the show’s message is consistent: AI is becoming more useful when it is embedded in real context, connected to action, and designed around human needs rather than abstract capability demos.