Glenn Solomon

The agents in your support queue might be the biggest predictor of what you should build next.

This week, I sat down with @rauchg at ICON Innovation Week. We covered a lot of ground, but the moment that perhaps best captures Guillermo as a CEO came when he mentioned he still takes regular rotations with the @Vercel support team.

These days, nearly all of Vercel support tickets can be managed by their own internally-developed agent, so the problems escalated to humans are truly novel. In his most recent support rotation, a question came up where an agent was requesting human confirmation before taking action on a high-risk task. Is this a hint toward the future and how humans and agents will interact?

Guillermo has an outsized ability to stay extremely close to the developer, and now, the agent.

Looking forward to continuing the conversation with G next week at @TomorrowXSummit. What questions should I ask him?

Are too many AI startups selling picks and shovels and not enough mining the gold, as some say? I see it differently.

There are plenty of companies mining for gold. It’s just hard to win doing it. A model release, a feature a lab ships for free, a competitor built on the same APIs over a weekend. App franchises can be threatened very quickly today.

Embedded infrastructure holds up better for 2 reasons:

1 is old: Once a team’s deploys run through @Vercel, or its infrastructure is defined in Terraform, replacing it means rewriting the entire system around it.

2 is new: @rauchg’s warning is that you can now burn unlimited tokens on optimizations no user will ever notice.

Competitors have your models. They don’t have your traces, nor your data.

Great to have @mitchellh onstage at our 2026 @notablecap AGM to hear about what he’s building at @superlogical. Just this month, the team released Rex for public beta testing - a drop-in replacement for any terminal a developer may be using today.

Since backing Superlogical’s seed round, it’s been thrilling to watch Mitchell and team back building 0 to 1. Few people can balance the product taste and finesse that developers crave with the relentless pace of shipping that early-stage AI startups require, but Mitchell and team can. Excited for more developers to get their hands on Rex and for the product roadmap to continue to unfurl.

In every platform shift I’ve seen, developers picked the winner before the enterprise did.

The @AnthropicAI S-1 shows the same pattern in AI spend. Claude Enterprise users who don’t touch Claude Code spend under $75 a month on average. The top 5% of enterprise users on Claude Code spend roughly 20x that.

A chat window answers a question. Claude Code does the work: a migration, a test suite, a pipeline someone used to maintain by hand. Spend follows the work, and right now the work runs through developers.

Follow the developer, and you’re serving the knowledge worker who comes next.

What actually costs money when an agent runs? Most of the bill is context: the agent re-reads every step.

@cursor_ai's breakdown:

-38% tool outputs the agent reads back
-17% system prompts & tool definitions
-10% reasoning

🔽 Token prices keep falling. 🔼 Tokens per task keep rising.

Architecture decides which wins. If your margins haven't turned, looking at what your agent reads may be more important than what you’re paying per token.