Coding Isn't About Code Anymore

Summary

Theo argues that AI is changing software engineering from primarily writing code into creating and maintaining an environment where code—and AI agents—can produce good results. He calls this work “gardening”: improving the codebase, tools, checks, and workflows so useful changes grow in the right direction and poor ones are caught before they spread.

From coding to gardening

• Theo says engineers have traditionally worked across layers of abstractions, assembling components and choosing where to draw boundaries. AI can take over more of the direct implementation, but that does not eliminate the engineering challenge.
• The work shifts toward building a healthy “garden”: removing sources of friction, providing useful guidance and feedback, and monitoring changes for regressions or low-quality output.
• This maintenance matters especially with AI because agents can produce code quickly, including “slop.” If the codebase and development process are unhealthy, problems accumulate faster than teams can fix them.

Why he values platform work

Theo traces this perspective to his time at Twitch, where he saw how product teams depended on platform teams and how changes to shared tools could help many engineers work more effectively.

• He describes moving from backend work into React and TypeScript, and becoming interested in how a well-understood system makes it easier to reason about changes.
• A formative example was a colleague improving documentation and confusing code after Theo’s PR revealed that both had led him astray. Theo took this as a lesson to address the source of recurring mistakes, not just fix individual symptoms.
• He emphasizes that product and platform work should inform each other. Platform improvements make it easier to build products; product work, in turn, reveals where platforms need to improve.
• At Twitch, he sometimes helped bridge teams when a product team’s assumptions about platform constraints were wrong. In one example, a team’s complicated approach to representing dashboard layouts was unnecessary: the GraphQL platform lead agreed that storing a JSON blob was the simpler solution.

Building with fewer resources

After leaving Twitch, Theo initially tried to recreate the infrastructure and team capabilities he had relied on, but found that maintaining all of it himself was too much.

• He created the T3 Stack by combining existing tools: Next.js, TypeScript, Tailwind, Prisma, and tRPC. He highlights tRPC’s ability to provide a TypeScript contract between the backend and frontend without defining a separate API schema.
• He describes regularly alternating between product work and platform improvements: when a feature is harder to build than it should be, improve the underlying setup, then return to the feature.
• He presents this balance as a way to make limited development time more productive, not simply as a pursuit of faster coding.

Examples of platform improvements

• T3 Chat’s migration to Convex: Theo replaced a custom synchronization engine with Convex, which he describes as combining a backend, database, API layer, and client/server SDK. The migration involved a large, complex change that users mostly did not notice, but it made the system easier for teammates to work on and improved cross-device synchronization. Theo says the simpler setup let the team continue shipping while he was unable to code for a period.
• TypeScript performance: At Twitch, slow type-checking and development startup times became a major source of friction. Theo recounts helping connect Twitch’s browser-platform team with the TypeScript team to discuss performance.
• Large migrations: He compares a Twitch-era migration involving thousands of files and manual coordination with what he says can now be handled with a single AI prompt.
• T3 Code testing: Theo says he removed about 250,000 lines of code by deleting roughly half the unit tests—around 11,000 tests in total—because many were unhelpful and generated unnecessary edge cases. He says only four tests had to be restored and that the deletions had not caused regressions that those tests would have caught. He also describes using repository guidance and investigating why contributors’ agents produced bad tests, with the aim of fixing the underlying cause.
• Regression monitoring: To prevent performance regressions in T3 Chat’s thread-transfer protocol, he built a CI check that simulates real thread data and measures how much data passes through the transfer process. The check runs on pull requests, flags meaningful regressions, and gives agents explicit feedback about a performance goal. Theo says it takes less than about 45 seconds to run.

Implications for AI-assisted development

Theo argues that slow development servers, CI, tests, or release pipelines can constrain agents even more than they constrain human developers. Slow feedback also limits parallel work and can waste agent context or tokens. He therefore sees continuous platform maintenance as central to effective AI-assisted development.

He says that keeping a codebase healthy can increase the rate at which a team ships changes. As an example, he claims T3 Code’s pull requests grew from roughly 2–20 merged PRs per day to about 60 on a busy weekend day and more than 400 on active days. He attributes the increase to the team’s improved development environment and processes.

Sponsor segment

The video includes a CodeRabbit sponsorship for Rabbit Secure. Theo describes it as offering deep scans to find existing security issues and scans of incoming pull requests to help identify vulnerabilities and assess the reachability, exploitability, and potential blast radius of changes. He warns that AI can produce insecure code and that exploits may be attempted soon after a patch becomes public.

Review, guide, or tutorial

This is an analysis and personal argument, not a step-by-step tutorial or product review. It includes practical examples of platform work, testing, performance monitoring, and AI-agent workflow improvements.

Main speaker and sources

• Theo — the video’s primary speaker and source of the analysis, personal work history, and product examples.
• Referenced colleagues and teams — including Twitch’s browser clients, GraphQL, and product teams; the TypeScript team; and Convex collaborators.
• CodeRabbit — sponsor of the security-tool segment.
• Chat participants and a referenced social-media post — briefly cited in relation to the gardening analogy.