# The next era of AI coding

[Watch on tuber](https://tuber.guzus.xyz/video/8h9j2rskP14/summary) · [YouTube](https://youtu.be/8h9j2rskP14)

**Score: 59/100 · skim**

Cursor's opening keynote says software engineering is moving through three eras: tab autocomplete, single agents, and now "teams" of autonomous cloud agents. The speaker backs this with Cursor's own usage data: at Cursor, 30% of internal PRs are now built end-to-end by agents, and about 75% of enterprise customers' code is AI-generated. Engineers are becoming "agent managers" whose main jobs are delegating and reviewing, and Cursor's experiment with a fully autonomous agent team building a web browser shows how far this could go, though the speaker says it isn't production-ready.

- **0:04-1:38** — **Star Wars as a "before and after" moment.** The event is the first in Cursor's new office, a former movie theater where the original Star Wars premiered in the late 1970s. The speaker describes the Dykstraflex, a computer-controlled camera that could repeat the exact same moves so model shots could be filmed in layers and combined. It's offered as an example of technology redefining an industry, and he says software engineering is at a similar turning point now.

- **1:38-2:30** — **How software used to be built.** Until a few years ago, the only way to build software was to type logic in formal programming languages. He shows a simple Python script that builds a list, loops over it, and computes values. He notes that many leaders, and increasingly their engineers, rarely look at code like this anymore.

- **2:30-3:10** — **Why traditional coding is slow and expensive.** Engineers have to turn high-level ideas into tiny, exact steps: if/then branches, loops, moving data from one place to another. A feature a human could describe in a couple of paragraphs can take tens of thousands of lines of code. That verbosity is a big reason software eats hundreds of millions of dollars of R&D budget.

- **3:10-4:21** — **The hidden weight of existing code.** Professional engineers work in codebases of tens of millions of lines, so a change can touch thousands of files of existing logic. He compares software to a Gothic cathedral: the cathedral shows how hard it was to build, but software hides its complexity behind simple buttons. That's why product managers and non-technical staff have long been frustrated by how slowly features ship.

- **4:21-4:58** — **The data behind the shift from tab to agents.** In 2025, agent requests in Cursor grew more than 15x year over year. At the start of 2025, tab-completion accepts outnumbered agent requests about 10 to 1; now agent requests are well ahead. This matters more than the raw counts suggest, because one agent request is usually a large chunk of work, while a tab accept is a few characters or lines.

- **4:58-5:29** — **The "teams era" inside Cursor.** About 30% of Cursor's own PRs are now written end-to-end by agents with no human intervention. These agents run on their own remote cloud computers and can work on a task for hours or even days. The speaker admits this is a cherry-picked, unusually forward-thinking company.

- **5:29-5:51** — **Enterprises are following.** In Cursor's enterprise segment, AI-generated code went from about 15-20% roughly a year ago to about 75% now. Humans in those companies increasingly don't write syntax themselves; they hand the work to agents. The point is that the trend isn't limited to startups or mid-market companies.

- **5:51-6:48** — **Engineers become agent managers.** He suggests a new mental model for an organization: alongside a few thousand human engineers, you now have tens of thousands of "ghost colleagues." Companies should pick the smartest agents that also give the best value for the price. They also have to rethink workflows, because the way humans work with AI, AI works with AI, and humans work with each other all change.

- **6:48-7:20** — **Review is the new bottleneck.** He warns that the move from tab to agents to teams can hurt an organization. It can produce code that can't be maintained, makes bad architecture decisions, and is full of bugs. To prevent that, engineers now spend much more time reviewing: reading the code, using built versions of the software, and testing.

- **7:20-7:56** — **More parallel work and more context switching.** In the agent era, agents run locally and edit the same copy of the codebase, so one person can babysit about one to three at a time. In the teams era, each agent has its own computer, so one person can oversee dozens of long-running tasks at once. The cost is more multitasking and context switching for the human.

- **7:56-9:32** — **The far future: an autonomous browser experiment.** Cursor tested how far agents can go with as little human involvement as possible. An agent team ran for a full week with no humans, opened many PRs, wrote about 3 million lines of code, and produced a "mostly functional" prototype web browser. Early on it couldn't render apple.com or a test landing page well; by the end of the week it rendered pages fairly well. Slides showed lines of code, how many agents ran at once, tokens used, and a time-lapse of the codebase being built. He repeats several times that this is very early, experimental, and "not ready for prime time." Its value so far is shaping Cursor's product decisions for the teams era. The talk ends here, with no separate outro.

**Practical takeaways:**
- **Treat engineering as delegation plus review.** Expect engineers to spend less time typing code and more time scoping tasks, managing parallel agents, and checking their output.
- **Invest in review capacity before scaling agents.** Code review, testing, and build checks keep agent-written code from piling up bad architecture and bugs.
- **Look at cloud-hosted agents for real parallelism.** Agents that run locally on one codebase cap a person at about 1-3 at a time; agents with their own environments make dozens of concurrent tasks possible.
- **Plan workflows for a mixed human and AI workforce.** Choose agents on both quality and price-performance, and redesign how people and agents hand off and coordinate work.
- **Discount the headline numbers.** The 30% and 75% figures come from Cursor's own product data and its own team, and the autonomous browser is an openly experimental demo, not proof of production readiness.

## Audience signal

- No high-signal audience themes available.