You're Building AI Agents on Layers That Won't Exist in 18 Months. (What this Means for You)
TL;DR
The real AI battleground is moving below agents — Nate argues we’re watching a cloud-scale infrastructure shift from “human-first tools” to “agent-first primitives,” and the winners will be the companies building the system-call layer agents need to act in the world.
Today’s agent stack is not Lego bricks — it’s Legos mixed with wooden blocks — his core warning is that these tools are marketed as composable, but in practice they’re uneven, immature, and full of architectural bets that builders need to understand before wiring them together.
Compute and sandboxing is the most mature layer, but even here the key choice is persistence vs. ephemerality — E2B (~$32M, Firecracker microVMs) treats sandboxes as disposable, while Daytona ($24M Series A, 90ms cold starts) bets on persistent container-based agent sessions.
Identity, memory, and integrations are useful now but full of platform risk — AgentMail ($6M seed) makes email a pragmatic identity shim, Mem0 ($24M, 41,000+ GitHub stars, 14M downloads) treats memory as curated infrastructure, and tools like Composio ($29M) solve the painful N×M integration mess.
Stripe’s new provisioning layer points to a trust and billing stack built for agents, not humans — Nate highlights Stripe Projects as the first credible way for agents to provision services and pay securely, with database startup times around 350ms and tokenized payment credentials kept in Stripe’s vault.
The biggest open opportunity is orchestration, and that’s where the next infrastructure giant may emerge — he compares today’s multi-agent tooling to the pre-Kubernetes era and says whoever solves scheduling, coordination, supervision, failure recovery, and FinOps-grade observability for agents could own the most valuable position in the stack.
The Breakdown
This Isn’t About Agents — It’s About the Layer Underneath
Nate opens with a big claim: a new AI infrastructure stack is being assembled in public, backed by billions, and most people are staring at the wrong layer. He frames it as the third major platform shift after cloud (2006–2010) and microservices (2012–2016), except this time the customer for infrastructure is the agent itself.
The “Lego Bricks” Pitch Is Mostly Marketing Right Now
He warns against the tidy fantasy that agent infrastructure is already modular and composable. His metaphor is great: this isn’t a box of matching Lego bricks — it’s Legos and wooden blocks all being sold as Legos, and nobody can tell what actually snaps together.
Compute and Sandboxing: The Most Real Part of the Stack
The first layer is where agents safely run code, and Nate says this is already fairly mature. He contrasts E2B’s Firecracker microVM approach and roughly $32 million in funding with Daytona’s Docker-based architecture, $24 million Series A, and claimed 90-millisecond cold starts, then draws out the real philosophical split: should agent sessions be disposable or persistent?
Identity and Communication: Email Works, but Maybe Only as a Shim
Next he moves to the weirder layer: how an agent exists on the internet and proves who it is. AgentMail, which raised a $6 million seed from General Catalyst with angels like Paul Graham and HubSpot CTO Dharmesh Shah, embodies the current pragmatic answer — give the agent an email address — but Nate keeps stressing that email may just be a cockroach-like survivor, not the native long-term protocol agents actually deserve.
Memory Is Valuable Now — and Vulnerable Later
On memory, he spotlights Mem0 as the leader: $24 million raised, 41,000-plus GitHub stars, 14 million downloads, and AWS selecting it as the exclusive memory provider for its agent SDK. What he likes is the framing that memory is active curation, not chat transcript hoarding, but he also makes the platform-risk point bluntly: if OpenAI or Anthropic build good enough native memory, standalone memory startups could get crushed.
Tooling and Integrations: Solving the N×M Nightmare
The next layer is tool access, where agents need to work with Slack, Jira, Salesforce, GitHub, Google Workspace, and everything else enterprises already use. Nate calls out Composio and its $29 million in funding for handling auth, connectors, observability, and the ugly error-prone middleware problem that otherwise turns every agent deployment into an unsustainable combinatorial mess.
Stripe Signals a New Trust Layer for Agent Spending
Provisioning and billing is the newest part of the stack, and Nate sees Stripe Projects as the first serious infrastructure for agent-to-service transactions. His point is that agents could already do almost everything except the final trust step — account creation, provisioning, and payment — and Stripe’s product closes that gap with tokenized credentials, terminal-first workflows, and databases ready in about 350 milliseconds.
Orchestration Is the Kubernetes-Sized Prize
He saves the biggest opportunity for last: coordination across many agents. Citing Gartner’s reported 1,445% surge in multi-agent system inquiries, he says current tooling is still framework-level, not infrastructure-grade, and lays out what’s missing — scheduling, merge/conflict handling, supervision hierarchies, financial observability, and standard recovery patterns — before arguing this is where the next defining infrastructure company will be built.
What Builders Need to Internalize Right Now
Nate closes with three practical warnings: reliability compounds in the wrong direction when you stitch together too many immature primitives, transitional lock-in is real when you build on shims like email, and “agent sprawl” is going to echo the worst excesses of microservices. His bottom line is stack literacy: whether you’re a founder, IC, or business leader, you need to know which of these six layers you’re actually betting your business on.