AI Agents Are Coming, but the Real Trade Is Elsewhere

AI agents are coming, but the money is not in the agents themselves. In the October 5, 2026 episode of Milk Road Stocks, host LG and researcher Vincent map out layer by layer which companies the era of personal chief-of-staff agents will enrich and which it will erase. LG opens with a dishwasher story: researching new models with Meta's agent Muse returned tidier results than ChatGPT and Claude.

At the top of Vincent's map sits the human, directly below a chief-of-staff agent designed as the single point of contact. Juggling 15 separate agents and 10 open chat windows at once is a job nobody can do; the fix is telling one door everything from a Paris trip to a stock purchase. The host likens the idea to the iPhone: just as we run everything from banking to music through one device, we will soon tell one agent to do every job.

The request travels downward from the chief-of-staff agent: first through identity and security filters verifying the agent truly belongs to us, then to the SaaS platforms' own agents that do the work, and finally to the payment at the bottom. In this design the human leaves the loop entirely; an agent acting for us deals with agents on the other side. For Vincent the investment question starts here: which layer of this stack will accumulate value, and which will evaporate?

Commoditization of the agent layer

The topmost agentic layer , the user-facing tier of Muse- and OpenAI-style agents, is Vincent's least favorite place. His case is threefold: brutal competition, enormous compute costs, and abundance constantly pushing prices down; he expects this tier to become a commodity like oil and gas, drifting toward zero pricing. According to Storyboard18, citing Sensor Tower data, Muse reached 5 million US downloads in just 22 days, well ahead of the 56 days ChatGPT needed for the same milestone, and passed 3 million weekly active users.

The second blow comes from portable context: because the chief-of-staff agent knows its user inside out, swapping the engine underneath becomes nearly costless. Once every agent clears a good-enough quality bar, each substitutes for the other and customers always migrate to the cheapest provider. Vincent argues that after quality stabilizes, price is the only differentiator left, permanently grinding down margins.

The third blow lands on advertising: AI agents are not suggestible creatures swayed by subjective ads but rational buyers coldly comparing price and specs. If the shoe ad you saw on Instagram sends your agent hunting and it returns five equally good alternatives at lower prices, the ad's value collapses to zero. LG pushes back with Netflix: subsidize the market, capture it, then sell subscriptions, a model that could work here too; Vincent replies the ad pie will not grow in an agent economy, only shrink.

Identity, security and the SaaS split

The first layer Vincent considers solid for the long run is identity and security: how will a platform know the agent in front of it is truly LG's agent, and what happens when malicious agents slip into the system? Okta, CrowdStrike and Palo Alto all play exactly this problem; CrowdStrike's Agentic Identity Provider, unveiled at Fal.Con in September 2026, promises every AI agent a trusted identity with only time-boxed, least-privilege access to enterprise systems. The catch is price: this story is already written into the shares, so Vincent keeps these names on a watchlist as expensive.

For all the year's 'SaaS is dead' headlines, Vincent splits the picture in two: software earning money purely on slick interfaces and per-seat licensing is genuinely under threat. According to the Gartner analysis published on CIO, agents put 234 billion dollars of enterprise software spending at risk by 2030 while rendering per-seat licensing obsolete, because agents never look at the interface. That is exactly what Gartner calls 'agentic arbitrage': as agents instead of humans do the work, the premium once paid for usability evaporates.

The other half is Vincent's favorite: backend platforms giving agents data, context and workflow triggers. Structures like Robinhood, Palantir, ServiceNow and Salesforce, which can actually execute a travel booking, a stock purchase or a money transfer, will serve a new customer class replacing the human buyer. The Airtable example under Bending Spoons shows winners open their platform to agents while building their own agent layer on top. In Vincent's phrase: agents are not killing SaaS, they are changing SaaS's customer.

The end of interfaces and payment rails

On the consumer side the picture is harsher: Vincent and LG agree that hopping site to site hunting for logins and downloading a separate app per chore is already unbearable. Once 95 percent of traffic comes from personal agents, every interface from airline pages to commerce checkouts will be designed for agents, not people. That is an open bear thesis on the Apple App Store model beyond 2030; on the enterprise side, hybrid screens where humans supervise agents survive because the cost of error is high.

The layer Vincent calls 'least noticed but most loved' is payments: agents will not use credit cards or bank wires but pay pennies in stablecoin per API call. According to x402 protocol figures compiled by CryptoBriefing, agents executed more than 23 million stablecoin transfers in the last 30 days, with the average payment worth only about 30 cents; totals reached 75 million transactions and 24 million dollars in volume. Bitwise's 2025 NEAR report and Goldman Sachs micropayments charts tell the same story: machine transactions dwarfing human counts demand cheap, scalable rails.

To the timing question Vincent gives a three-stage answer: agent infrastructure settled in recent weeks, regulation trails behind, and scale erupts in 2027-2028. According to Reuters, the SEC issued a five-year Innovation Exemption for tokenized stock trading, filling with its own rules the gap left by the stalled Clarity Act in the Senate. Believing the market has not priced this future yet, Vincent frames today as 'buying Micron and Bloom Energy in Q4 2025': back then nobody talked about memory and energy bottlenecks.

Compute: the layer with no ceiling

The finale is the layer with the broadest consensus: in this world compute demand has no ceiling. According to a16z charts reported by Tomshardware, agents already burn five times more tokens than humans on OpenRouter, and that ratio is expected to climb toward tenfold; agent token usage is up 14-fold since the February 2026 crossover. Vincent's analogy is memorable: what water, food and sleep are to humans, memory and power are to agents; trillions of agents against 9 billion people. The Pro portfolio's MU, BE, HOOD and CRDO bottleneck names are this thesis expressed as positions.