Wei Sun
CTO at Arcade · Palo Alto
Building with AI agents, and helping teams make them stick.
I lead engineering at Arcade and ship product daily with parallel coding agents, cross-model review, and tests as the contract. I also help engineering and non-engineering teams turn AI tools into a way of working, rather than a pile of seats.
Now
Updated October 2026
- Leading engineering at Arcade, shipping product every day with the agent workflows described on this page.
- Running several coding agents in parallel (Claude Code and Codex), each in its own git worktree through Paseo or Conductor.
- Bringing agents to non-engineering teams: internal tools built in Replit, and Hermes Agent for recurring, multi-step work.
Approach
What makes AI agents pay off
Buying the tools is the easy part. The gains come from how work is split, reviewed, and merged. These are the rules I run my own engineering on.
-
Parallel, isolated agents
One engineer can drive several workstreams at once when each agent has its own worktree, branch, and dev server. Nothing collides, and every change is reviewable on its own.
-
One model writes, another reviews
A second model reviewing the first catches what the author model is blind to. CodeRabbit sits in the pull request, and a human still owns the merge.
-
Tests are the contract
Agents move fast against a spec they can check for themselves. Without tests you end up reviewing intentions, and review becomes the bottleneck again.
-
The operating model outlasts the tools
The tools turn over every few months. Conventions, permissions, review policy, and clear ownership are what keep paying off.
-
Efficiency is designed, not rationed
Token spend follows how the work is set up, not how much people hold back. Match the model to the task, keep context focused, and scope work so agents land it cleanly. In the teams I've worked with, that has never needed more than $200–300 per engineer per month, with engineers running full tilt and never limited.
How it's measured
The 3× is merged pull requests across every repository Arcade's engineering team owns, January to September 2026 against the same months of 2025, after we rebuilt how we work around coding agents. Release and deploy PRs are excluded, because they re-attribute the whole team's work to whoever cut the release. PR counts measure volume, not difficulty or value, so read the figure as a trend rather than a productivity score. The $200–300 spend figure covers the same period.
Toolkit
What I reach for today
Tools I use myself, chosen per team. This list changes as the tools do; the approach above doesn't.
- Agentic coding Claude Code · Codex
- Agents that read, edit, run, and test code in the real repository, not a sandbox copy of it.
- Parallel orchestration Paseo · Conductor
- Each agent in its own git worktree, so several streams of work run side by side without stepping on each other.
- Review CodeRabbit · Codex
- AI review in the pull-request flow, alongside a second-model pass and a human merge bar.
- Beyond engineering Replit · Hermes Agent
- Replit when a team needs a working internal tool without setting up a dev environment. Hermes Agent, the open-source agent from Nous Research, for persistent, scheduled, multi-step work.
Work together
Hands-on help, then your team runs it
Alongside my role at Arcade, I work directly with teams on this. Every engagement starts with a 30-minute call, no prep needed, to map where your team loses time, and ends with your team running the system without me.
Engineering teams
For CTOs and VPs of Engineering at Series A–C companies, roughly 10 to 80 engineers.
You have AI seats, but delivery looks the same. I help you set up parallel agents, cross-model review, and the conventions that make agent-written code mergeable, without a runaway token bill.
Non-engineering teams
For operations and finance leads.
Operations, finance, and other business teams with work that is repetitive but still needs judgment. I help you build the tool or agent yourselves, and keep it running.
Ways to work together
-
Agentic Engineering Audit About a week, fixed scope
I review how your team uses AI today and write up what to change, in what order, and what it should cost to run. Useful on its own, and the plan for everything below.
-
Team workshop On your codebase
Hands-on sessions on your real repositories and workflows, not a demo project.
-
Rollout 2–4 weeks
I configure the tools against your real repositories and workflows, train the people who will use them, and hand off a written operating playbook.
-
Advisory Month to month, optional
Refining review policy, agent conventions, and tool choices as the stack evolves. Only if it makes sense for your team.
Probably not a fit if
- You want an AI strategy deck rather than working changes to how your team ships.
- You want someone to run your agents for you indefinitely. The goal is that your team runs them.
Board seats
I'm available to join company boards. I bring a practicing CTO's view of how AI is changing what a company can build, what it costs to build it, and where the technical risk sits.
Email about a board seat →About
About Wei
I'm CTO at Arcade. Before that I was CTO and then CEO of Upduo, acquired by Arcade in 2025, and earlier worked on machine learning at Apple. I studied at MIT.
I'm CTO at Arcade, where I lead engineering and ship product daily with the same AI-native workflows I help other teams adopt. Before that, I was CTO and then CEO of Upduo, which was acquired by Arcade in 2025. Earlier in my career, I worked on machine learning at Apple.
I studied at MIT and have spent my career at the intersection of engineering leadership and applied AI. When agentic coding tools started getting good, I didn't delegate the evaluation. I rebuilt my own workflows from scratch, kept what worked, and turned it into repeatable playbooks.
- Current
- CTO, Arcade
- Previously
- CTO & CEO, Upduo (acquired by Arcade, 2025)
- Earlier
- Machine Learning, Apple
- Education
- MIT
If your team has AI seats but works the same way it did before, let's talk.
Email a few lines about your team and where work feels slow. I'll reply with an honest read on whether I can help, typically within one business day.