Shawn Wang on MCP, AI agent horizons, and why building infrastructure companies is a warning sign

Summary

Shawn Wang — known as Swyx, host of the Latent Space podcast and organizer of the AI Engineer Conference — argues that the AI industry is spending too much time on infrastructure and not enough on the hard problem of building reliable agents.

MCP

MCP, the integration protocol released by two Anthropic engineers in London, solves what developers call an M×N problem: every app historically had to write its own integrations from scratch, so switching apps meant rewriting everything. MCP collapses that into an M+N problem where each integration is written once and reused across any compatible agent. Wang's framing is deliberately deflating — MCP is a protocol, not a product, and no one got excited when REST was invented. The excitement comes from what REST enabled, and MCP's payoff will be similar: agents that can do far more out of the box, without users waiting months for their favorite app to ship a custom integration.

Wang does flag one genuinely bullish technical detail: MCP servers can also act as clients, meaning they can spin up and orchestrate fleets of other MCP agents without the user knowing. That capability makes the whole ecosystem substantially more agentic. He expects it to surface meaningfully by end of 2025 or early 2026.

On durability, Wang is not worried about MCP being steamrolled by better models — more context window capacity actually makes MCP more useful, not less. The only credible challenger, in his view, is Google, which already has native first-party integrations across Gmail, Calendar, and YouTube. But Google lost the frontend standards war with Angular versus React, and Wang thinks they're unlikely to mount a serious fight here.

For startups trying to build MCP registries, the headwind is that Anthropic is building its own. Wang's position is blunt: the burden of proof sits with the independents, not with the lab's official solution.

Agent infrastructure glut

At YC's most recent demo day, there were more AI agent infrastructure companies than agents Wang has actually tried to use. His read: building AI agent infrastructure is what developers do when they have no other ideas. The real scarcity is reliable agents that solve concrete problems, and more picks-and-shovels companies don't fix that.

Agent horizons

METER (Wang approximates the acronym) published research tracing agentic capability back to 2019 and found that the task horizon — how long the best available model can operate autonomously at 50th-percentile human capability — is doubling every three to seven months. As of the measurement, that horizon sits at roughly one hour, with Claude 3.7 Sonnet as the benchmark model. Extrapolating forward gives a rough schedule for when one-day, one-week, and one-month autonomy become plausible, which Wang suggests is a useful clock for deciding what products or companies to build toward.

Cursor's trajectory — zero to $200 million ARR in roughly two years — is his clearest example of what incremental, clear-eyed progress on code generation can produce, moving from single-line autocomplete to full app and PR generation.

AI 2027 and scaling

On the AI 2027 report, Wang credits the authors with doing genuine public service by drawing trend lines that most practitioners are too day-to-day to see. His caution is standard but worth stating: every exponential eventually hits a sigmoid. COVID case projections in April 2020 looked like they'd exceed the global population; they didn't, because of invisible limiting factors that only appear when you reach them. The 2027 predictions on coding agents and robotics he finds credible; the bioweapons and political sections he considers much more speculative.

Apple and the OpenAI phone

Apple's AI fumble is, in Wang's view, not a frontier problem — it's a deployment problem. The models exist. Apple hasn't shipped. Wang is more interested in OpenAI competing with Apple than with Google. ChatGPT is already the sixth most-visited website in the world, and OpenAI is essentially running the Google playbook. The more interesting contest is hardware. Wang believes an OpenAI phone — which he notes is confirmed to be in discussion with Jony Ive — would be the first serious challenge to the iPhone since Steve Jobs introduced it, and that it would force Apple to finally get its AI integration in order. Consumer appetite for AI hardware is already there; what's been missing is a well-resourced lab willing to build it seriously.