The phrase "email is primitive" could sound like the setup to a pitch for yet another messaging startup promising to kill the inbox. But when Ethan Byrd says it, he means something different—something closer to foundational. After logging time at Microsoft, AWS, Facebook, and Google, then serving as CTO of Actual AI (which raised $3.2 million in September 2025), Byrd has spent enough years watching enterprises fumble with autonomous systems to recognize a pattern. The question isn't whether AI agents will proliferate. It's how they'll actually talk to each other—and to us—without everything breaking.
His answer, delivered through his startup Primitive (part of YC's Spring 2026 batch), manages to be both radical and almost painfully obvious: Use email.
Yes, email. The technology most founders treat as legacy infrastructure, something to "disrupt" rather than embrace. Byrd's contrarian thesis is that the humble email protocol, with all its decades-old quirks, might be exactly what the agent economy needs to move from pilot programs to production scale.
The Gap Between Hype and Reality
Market projections for AI agents have that familiar venture capital smell—big numbers getting bigger. Grand View Research pegged the agents market at $7.63 billion in 2025, growing to $10.91 billion by 2026, with forecasts stretching to $182.97 billion by 2033. Gartner, in August 2025, predicted that 40 percent of enterprise applications would feature task-specific agents by late 2026, up from under 5 percent.
But step back from the hockey sticks and a messier picture emerges. Deloitte's State of AI report found that while nearly three-quarters of organizations plan to deploy agentic AI within a couple years, only about one in five claim to have mature governance frameworks ready. McKinsey's data from late last year showed 23 percent of organizations actually scaling agentic AI in at least one function, with 39 percent still tinkering.
That gulf between experimentation and production? It's not just about model quality or accuracy benchmarks. It's about the unglamorous stuff: authentication, error handling, audit trails. How do you know what your agents actually did? How do you stop them when they start doing something wrong?
As Primitive's YC directory page puts it, the focus is on "communication for agents." Which sounds simple until you try scaling it.
The Protocol Wars
The timing of Primitive's bet on email gets more interesting when you consider what else is happening in agent communication. Anthropic introduced the Model Context Protocol (MCP) in November 2024, focused on standardized agent-to-tool interaction, now governed by the Linux Foundation's Agentic AI Foundation. GitHub, JetBrains, ServiceNow, Zendesk—they've all added MCP support over the past year or so. Google countered with its Agent-to-Agent (A2A) protocol in April 2025, focusing on agent-to-agent messaging, also under Linux Foundation governance, promising better interoperability.
These are sophisticated technical solutions. They're also, in Byrd's implicit argument, solving the wrong problem first.
Email has threading—those In-Reply-To and References headers that create conversation structure. It has authentication mechanisms (DMARC, SPF, DKIM) that major providers spent decades hardening. It has deliverability infrastructure that, for all its frustrations, actually works at scale. And perhaps most importantly for enterprises, it has a governance model that compliance teams already understand.
When an agent needs to loop in a human for approval, or leave a paper trail for audit, email's structure maps naturally to existing requirements. The primitive.dev documentation, refreshed this spring, reveals an infrastructure play built around these realities: SDKs in Node, Python, and Go; webhooks for incoming messages; proper threading; and what the docs call deliverability "gates" and recipient policies. Those last features hint at something the newer protocols haven't quite addressed yet—the governance layer enterprises actually need to put agents into production.
Primitive's documentation suggests customizable infrastructure with organization-specific configurations, indicating they're handling operational complexity while giving customers control. It's infrastructure work, the kind that doesn't make for flashy demos but determines whether systems actually run.
The Compliance Bottleneck
Here's where things get concrete. The FCC's STIR/SHAKEN requirements for caller ID authentication and the 10DLC A2P messaging registration process create real deployment bottlenecks for voice-based agents, requiring compliance from voice-based agent deployers. An April discussion thread among AI practitioners flagged this directly: "10DLC vetting delays and STIR/SHAKEN attestation can constrain deployment speed for phone/SMS-based agents." We're talking weeks or months.
The EU AI Act provisions began taking effect in August 2024, with general-purpose AI model obligations kicking in about a year later. ISO/IEC 42001, the first certifiable AI management standard, received its initial UKAS accreditation for certification in January 2026. As procurement teams start requiring these certifications—and they will—startups building agent infrastructure face an uncomfortable choice: build compliance in from day one, or retrofit it later when it's exponentially harder.
Email sidesteps some of this morass by riding on established standards. The infrastructure for managing domains, handling bounces, respecting unsubscribe requests? Already solved problems. Not exciting problems, but solved ones.
What Primitive Actually Is
The company, with a three-person team as of this spring, remains early-stage. No funding amounts have been disclosed publicly beyond the YC backing. They're hiring founding engineers with equity grants that signal both ambition and urgency: founding infrastructure engineers at $240K to $320K with 2 to 4 percent equity; founding engineers at $220K to $280K with 1.5 to 2.5 percent.
The product surface, visible through primitive.dev, emphasizes developer experience. The quickstart docs walk through sending and receiving emails via API, setting up webhooks, managing threading. That "outbound gating" feature isn't just nice-to-have—it's designed to prevent the kind of mistakes that could get an entire domain blacklisted, a genuine risk when autonomous agents start sending email at scale.
The positioning is careful, maybe deliberately so. YC frames the company as "communication for agents," not specifically email infrastructure. That leaves room to expand into other channels as the market evolves. But the immediate bet is clear: email will be the first channel where agents need production-grade infrastructure, and most companies don't want to build that themselves.
The Broader Landscape
Primitive isn't alone in carving out specialist territory within agent infrastructure. Oasis Security raised $120 million in a Series B this past March from Craft Ventures, Sequoia, and Accel, focusing on identity and access management for non-human identities. Other startups are tackling control planes (Aion), execution environments (Dynamis, Cowboy), and capability discovery (Anything.network).
The orchestration layer has consolidated around a few major players. Microsoft shipped Agent Framework 1.0 in April 2026, unifying Semantic Kernel and AutoGen. LangChain's LangGraph Cloud, initially launched in 2024, has become an actual production platform rather than a toy. CrewAI continues active open-source development.
Voice infrastructure evolved quickly, perhaps too quickly. OpenAI's Realtime API, announced last August, added SIP support for phone calling alongside remote MCP servers. Twilio and Vonage rushed to integrate, positioning themselves as the communications layer for voice agents. Twilio's SIGNAL conference this past May showcased Conversation Relay, PCI-compliant voice workflows, and an Agent Connect SDK.
But each of these layers assumes a maturity most enterprises simply don't have yet. ServiceNow's AI Control Tower, announced in May, attempts to provide observability and governance across MCP transactions. Even with enterprise-grade tools, though, the fundamental communication primitives remain more complex than they should be.
McKinsey's analysis from this spring emphasized that "infrastructure becomes backbone for orchestration and value capture" in agentic systems, stressing governance, observability, interoperability. Forrester's recent predictions pointed to a market correction focused on ROI and operational governance rather than continued hype cycles.
The Execution Challenge
Primitive's focus on email communication infrastructure might prove prescient—or it might be too narrow. While MCP and A2A solve real technical problems, they don't directly address human-in-the-loop workflows, approval chains, and audit requirements that enterprises need for autonomous agents to move from pilot to production. Email, with all its historical baggage and peculiar SMTP quirks, already has answers to these problems. Whether those answers are good enough is another question.
The challenge is execution. Building reliable email infrastructure that scales with the unpredictable traffic patterns of autonomous agents, maintains deliverability across major providers (Gmail, Outlook, the enterprise email security layer companies), and provides the governance hooks enterprises require isn't trivial engineering. With a small team and early-stage resources, Primitive will need to move quickly as larger players like Twilio and Vonage expand their own agent offerings.
Then again, there's value in focus. Perhaps doing one communication primitive—just email—exceptionally well for agents is the right play. As enterprises deploy the hundreds or thousands of agents that analysts keep forecasting, they'll need reliable ways for those agents to communicate that don't require rearchitecting every workflow. Sometimes the most boring infrastructure becomes the most critical.
Sometimes primitive is exactly what's needed.
