On Built This Week, AI in developer tools has two sides. One is what we use ourselves: Jordan said in September 2025 that his code was almost 100% written by bots, and our tools changed several times in a year. The other is founders building around coding agents: Amar Goel of Bito reviews code in large enterprise codebases, Adir Ben-Yehuda of Autonomy AI lets product people ship pull requests, and David Petrou of Continua manages agents and the developers watching them.
But when I start to get into my existing kinda gnarly code base that's been built for five, ten, twenty years, and I have millions of lines of code and hundreds of repos and different services, all of a sudden, tools start to kind of break down.
— Amar Goel, Episode 23
How companies we talked to use AI in developer tools
Bito: code review that understands the whole codebase
Bito indexes symbols and builds abstract syntax trees for a team's code, then reviews pull requests in GitHub, GitLab and Bitbucket or inside the IDE. Amar said PRs merge about 10 times faster, from fifty hours to five to seven, and reverted PRs drop by about 55%. Teams pick a noisy mode that teaches junior developers or an "essential" mode for production issues. He was blunt that a small vibe-coded project probably doesn't need it. (Episode 23)
Autonomy AI: product managers sending pull requests
Adir calls Claude Code and Codex the MS-DOS of coding and says Autonomy is building Windows. It analyzes an existing codebase, its standards and components, so a PM or designer can request a feature in plain language. In the demo it added Kanban and Gantt views to a task page, redesigned it for mobile and prepared a GitHub pull request. Adir said it launched six months earlier with over 160 companies, sees 86% of a few thousand weekly PRs approved on the first try, and customers use less Figma and Jira. (Episode 43)
Continua: an operating system for coding agents
David spent almost 18 years at Google. His product, Wheelie, takes a stated problem, splits it up, decides how many agents to run and where, and decides when to interrupt a person, down to their time zone. He argues human attention is now the scarce resource, and that Git, GitHub and React were designed for people, not agents. Wheelie was still on a waitlist when we spoke. (Episode 51)
Espresso AI: how an infrastructure team codes with AI
Ben Lerner uses Codex for dashboards and throwaway pandas scripts and says he hasn't visited Stack Overflow in over a year. Front-end tasks go to the on-call rotation instead of a dedicated hire. He put the gain at ten to thirty percent per engineer, and said Claude Code costs far less than a third of an engineer. (Episode 27)
What we built
- Strudel AI DJ mixer: Jordan built a two-deck mixer with an AI button that rewrites each channel's beat, mostly with OpenAI Codex. He had found Codex "highly underdeveloped" months earlier; over Labor Day weekend 2025 it built most of the tool.
- Bito Slack bot prototype: a Slack-style bot made in Google AI Studio that shows your last PR review, plus jokes, haikus and a disco mode. Amar said customers have asked for Slack alerts when reviews finish.
- Internal web app redesign via Claude Design handoff to Claude Code: Sam exported a redesign from Claude Design and gave it to Claude Code, which had the whole app live in production in about thirty minutes.
- RADAR: Matias Singer, a program manager at Ryz Labs who is not a developer, built it with Claude and Codex. It turns stand-up transcripts into suggested Jira tickets that he reviews before creating, which cut about four hours of work to minutes.
Our own stack kept moving. In Episode 11 Jordan said the team was switching from Claude Code to Codex. In Episode 40 he said our most advanced teams had moved on from Cursor, while many enterprises were only starting to adopt it.
What's still hard
- Nobody knows the code. Amar said PRs are so big that review takes days, and when a bug reaches production, the developer often didn't write the code.
- Serious backend work. Ben won't run prompted code against tens of millions of dollars of Snowflake workloads, and still writes a lot by hand.
- Developer burnout. David said running two, four or more agent windows exhausts developers, and teams skip postmortems because rolling forward with a fix is faster.
- Tool churn. Jordan said betting on what our workflow would look like by the end of the year "would be crazy."
- The gap for non-engineers. Jordan said building still means back ends, front ends and GitHub, which is a lot for a layperson. Adir added that most coding tools rewrite components instead of reusing them, which is where non-technical users get stuck.
Token costs are continuing to go down, but what's not improving is the availability of, you know, the human attention span.
— David Petrou, Episode 51
Episodes on AI in developer tools
- Episode 11: the Strudel mixer and why the team moved to Codex (September 2025).
- Episode 23 with Amar Goel of Bito: AI code review (December 2025).
- Episode 27 with Ben Lerner of Espresso AI: where AI coding falls short (January 2026).
- Episode 39 with Matias Singer: RADAR (April 2026).
- Episode 40: Claude Design to Claude Code, and the Cursor deal (April 2026).
- Episode 43 with Adir Ben-Yehuda of Autonomy AI (May 2026).
- Episode 51 with David Petrou of Continua (September 2026).
FAQ
What AI developer tools do you use?
We went from ChatGPT to Cursor to Claude Code to OpenAI Codex. In September 2025 Jordan said the team was moving to Codex, and in April 2026 he said our most advanced teams had moved on from Cursor.
Do AI coding tools work on large legacy codebases?
Bito CEO Amar Goel says they excel at new projects but start to break down on code built over five to twenty years, with millions of lines and hundreds of repos. Bito builds symbol indexes and syntax trees of the whole codebase to review code in that setting.
Does AI code review slow teams down?
Amar says the opposite: Bito customers merge PRs about 10 times faster, from about fifty hours to five to seven, because the agent gives feedback in minutes. Reverted PRs in the following month dropped by about 55%.
Can non-engineers ship code with AI?
Autonomy AI says product managers and designers can send pull requests on an existing codebase, with 86% approved on the first try. On our own team, a non-developer program manager built an internal Jira app with Claude and Codex in two months.