10 takeaways from AWS re:Invent 2025

1. Agents are no longer a punchline

Two years ago, "agentic" sounded like a joke. This year, it's the word on everyone's lips, but with receipts. Ajay Kulkarni, CEO & co-founder of Tiger Data, couldn't sleep after using Claude Code for the first time. Taylor Otwell, CEO & founder of Laravel, watches agents write better PHP than junior developers. Ronak Desai, CEO & co-founder of Ciroos, went from zero to enterprise production in nine months using AI-assisted development.

The shift isn't theoretical. Multiple founders told us 70-80% of code at their companies is now written by AI. Intercom's AI agent resolves 86% of customer conversations without human involvement. This isn't hype—it's happening.

2. The Claude Code moment

A surprising number of conversations included some version of "the first time I used Claude Code." It's become a watershed moment for technical founders, the night they stayed up until 4 AM, the feeling that they could build anything, the evangelical fervor the next morning.

What makes Claude Code different from ChatGPT isn't capability; it's agency. It writes code, commits to GitHub, runs tests, deploys. It does things rather than just answering questions. That distinction—from assistant to agent—is reshaping how people think about AI.

3. Documentation is the new moat

Here's an insight we didn't expect: documentation quality has become a competitive advantage. When LLMs write code for your platform, the quality depends on your docs. Vercel and Laravel have both invested heavily in AI-optimized documentation. Their frameworks get better AI-generated code because the training data is better.

This creates a network effect that favors incumbents. New frameworks face a cold-start problem, if there's no training data, the AI can't help developers use them. The rich get richer.

4. Pre-AI companies are having their moment

WorkOS, Vercel, Laravel, PlanetScale, incident.io—none of these are AI companies. They were built before the current wave. But they're perfectly positioned for it.

AI companies grow faster than any previous generation of startups. They need to be enterprise-ready earlier. They need authentication, authorization, observability, databases, and deployment infrastructure from day one. The "boring" infrastructure built over the past decade is exactly what AI companies need.

As Forrest Brazeal from Freeman & Forrest put it: "You can vibe code features. You cannot vibe trust."

5. The five-minute sweet spot

Cleric learned something counterintuitive: when they let their AI SRE investigate for 20 minutes before surfacing results, engineers felt disconnected. When they brought humans back at the five-minute mark, the experience transformed into collaboration.

This maps to a broader truth about AI agents. Long-horizon autonomous tasks don't work well without human grounding. Unlike coding (where you have tests and compilers), production debugging has no verification mechanism. You need the human.

Incident.io learned a similar lesson. Their first AI prototype was "utter garbage"—confidently wrong about root causes. It took 18 months of ground-truthing and evaluation infrastructure to build something that actually worked.

6. Open source models are winning

Baseten sees the gap between closed and open source models as perhaps two to three months, an eternity in AI time, but irrelevant for most industries. More importantly, the absolute capabilities of open source models keep crossing thresholds that unlock new use cases.

Philip from Baseten draws an analogy to operating systems: IBM, then Sun and Windows, then Linux everywhere. The same pattern is emerging with AI models, foundations from open source, with companies building on top.

7. Enterprise is pulling, not just being pushed

The surprise at re:Invent wasn't that startups are adopting AI. It's that enterprises with 30,000 employees are actively seeking out AI solutions. They have mandates to deploy AI. They're willing to work with nine-month-old companies.

This is a structural shift. Enterprise adoption of new technology is happening faster than ever before. The window for startups to land enterprise customers has blown wide open.

8. AI security has three fronts

Latacora splits AI security into three categories: AI of the company (securing AI features in your product), AI for the company (employees using ChatGPT), and AI against the company (attackers using AI).

That last category is underppreciated. The floor on phishing quality has risen dramatically. What used to be obvious scams are now sophisticated campaigns. Email security recommendations that worked five years ago don't cut it anymore.

9. Incidents are the new deploys

Incident.io's framing: incidents used to be quarterly catastrophes with clunky processes. Now they should be continuous, low-friction responses to anything that requires urgent attention.

This mirrors the shift from quarterly deployments to continuous deployment. When something is rare and scary, nobody gets practice. When it's continuous and routine, the process improves.

10. The 10x developer is becoming real

Multiple companies reported concrete data: developers using AI tools are significantly more productive. Vercel sees it in their logs—more commits, more deploys, faster iteration. This isn't a 10% improvement. It's multiples.

The 10x developer was always treated as mythical. Maybe it's becoming achievable for ordinary engineers with the right tools.

These takeaways are drawn from interviews conducted at AWS re:Invent 2025 with leaders from Baseten, Modal, Depot, PlanetScale, Vercel, Stedi, Intercom, Latacora, Acquired, incident.io, Laravel, Tiger Data, Ciroos, Freeman & Forrest, Browserbase, and Cleric.