DoorDash’s Andy Fang: Faster AI coding exposes the next bottlenecks

DoorDash co-founder Andy Fang says the company’s Cowork rollout aimed to raise AI fluency beyond engineering. Some employees were already coding with Claude Code, while others, including executives, still mainly saw AI as chat. He says connecting it to email, calendars or Slack showed people how it could help with knowledge work.

Boris Cherny asks how Fang began coding. Fang says he attended a coding camp at age nine because his mother had enrolled his brother. In 2013 he wrote DoorDash’s code in a Stanford dorm room, but as the company grew he stopped coding until Claude Code helped him ship production code again.

Fang usually codes in the terminal, sometimes trying the desktop app. He set himself a goal of not writing code manually and asks Claude Code to write it. He runs several sessions at once, using separate worktrees when they need to work on the same repository without conflicts.

His first attempt to ship production features without asking anyone for help failed because the agent could not configure his local environment. He says he repeated the experiment around the beginning of 2026 and it worked. Fang says he subsequently shipped production code in five languages, an experience that convinced him to keep trying the tools himself.

Cherny says an idea that fails with one model can work when retried with a newer one. Fang says DoorDash is encouraging teams not to become too attached to a working setup or agent: changing model capabilities may justify rethinking it.

Fang recommends that engineering managers set a goal of shipping actual production code, not just a local prototype, without asking someone else to do it. Going through the full process shows leaders what the models can do and what prevents engineers from moving faster.

He says DoorDash has told employees it will invest in tools and token budgets so they can explore coding agents. As managers try them, they begin to understand them too. Fang does not claim the rollout is perfect, but says people are moving along the learning curve.

Fang says he enjoys being able to prototype and build quickly, although running several agents and constantly switching context can be mentally taxing. Faster code production also creates new problems: employees notice how long merging takes, and continuous integration and delivery processes need rethinking.

DoorDash has invested in AI review agents to help people review code, Fang says. He also says more security issues are appearing across the industry, so teams need ways to use AI to catch them before production as throughput rises.

Fang says DoorDash introduced Claude Code in 2025 and worked with IT and security on a special procurement and review process so employees could get access quickly. He identifies an inflection point around late 2025 or early 2026 and says the company observed a correlation between adoption and team coding throughput.

He credits executive support and partnership among IT, security and engineering, along with not being overly strict about initial budgets. Cherny summarizes the approach as enabling exploration with guardrails; Fang says security is not negotiable and the security team remained involved in reviewing tools.

To make experimentation feel safe, DoorDash initially gave people tools and watched what early adopters found useful. It then encouraged people behind successful projects to advocate within their teams and produce written examples that others could share.

Fang says it is hard to enforce consistent sharing across a large organization, so teams need their own channels and forums. He wants people to share shortcomings as well as wins, including workflows that did not work and integrations that wasted tokens.

Fang says Claude Code and Claude models used through other tools have become widespread at DoorDash. To reduce friction around the resulting throughput, the company is investing in delivery infrastructure and AI code review.

DoorDash built an internal platform called Flux, with virtual machines in its cloud where employees can start security-approved Claude sessions with access to appropriate tools. Fang says its AI code review agent also runs on this platform, using the Agent SDK and Claude models.

DoorDash set goals for designers to ship code to production. Fang says the transition was not entirely smooth, but it pushed designers to participate directly in software development. He says many of the company’s best teams now embed designers and product managers in that cycle and use Claude to support them.

Fang challenged teams to complete projects in three to five times less time using AI. He says that constraint reveals obstacles in how people use coding agents and share standardized skills. In one example, he says, a single engineer completed a large migration in three weeks that historically might have taken four engineers a quarter.

Fang says other major product timelines have been cut by more than half. DoorDash is selecting teams for deeper experiments and asking successful teams to document what they did so others can learn from them.

Smaller teams can move faster because individuals own work across one or more domains and need less coordination, Fang says. Another lesson is to invest early in making a codebase friendly to agents.

One technical lead wrote more than 50 architectural principles in Markdown files in a repository so an agent could consult them while coding. A mobile team also standardized reusable skills for launching a simulator and testing workflows that would otherwise require tedious manual setup.

For user-facing features, Fang says, a code migration is not the whole job: teams must agree on designs and ensure they are solving the right product problem. Product, launch and design reviews had reasons for existing, but can become friction when engineering accelerates.

Fang says leaders across functions must reconsider how they work to realize the benefits of faster coding. For example, engineers could create a workable implementation and a designer could refine the details directly. He sees more self-sufficient individuals and teams as a way to move faster.

Fang suggests reimagining processes while recognizing that some serve important purposes, including compliance. He asks teams aiming for much shorter timelines to consider how agents could automate or speed up necessary steps.

DoorDash encourages mobile engineers to work more broadly across the stack, Fang says. Specialists are still needed for areas such as iOS architecture, memory, latency and performance, but fewer experts can record principles or build checks that help a larger group work across the codebase.

Cherny asks how to staff teams and give them room to experiment. Fang says he or another executive directly sponsors selected workstreams, helping teams raise and remove cross-functional obstacles. Simply asking people to experiment while keeping all the constraints of their normal jobs can cause them to give up.

Fang says the support includes token budgets and permission to bring non-coding barriers to leadership. DoorDash is still working out how to scale the approach; as examples succeed, it shares written accounts with other leaders so practices can spread.

Asked about return on investment, Fang says coding throughput is an imperfect and gameable metric but can be directionally useful. Faster completion with fewer people also frees the existing team for other work. For knowledge work, he says, DoorDash is earlier in the process of discovering valuable uses.

Fang offers sales teams preparing reviews for tens of thousands of merchants as a possible workflow to streamline. DoorDash is examining both time saved on common tasks and department-specific processes that agents could make substantially more efficient.

For engineering, the ultimate measure is whether customer value arrives sooner, not just whether code is merged faster, Fang says. For knowledge work, he sees value in improving the average employee’s productivity while continuing to explore domain-specific uses.

Cherny says a top-down automation effort can target the wrong workflow, whereas broad experimentation can uncover unexpected bottlenecks. Fang says DoorDash is identifying AI champions in sales, marketing, operations and support: people already experimenting who understand their own domain’s pain points.

Those champions can streamline a workflow with agents or create shared skills that help coworkers. Fang says naturally enthusiastic users tend to share their experiences with peers, making champions visible on both technical and non-technical teams.

For engineering leaders, Fang advises using AI themselves to understand its power and the anxiety that changes in work can create. He says some engineers identify closely with the craft of writing code, so leaders need empathy as workflows change.

He again recommends sharing written artifacts: they distribute lessons and can also be read by agents. To new graduates, Fang says to stay curious and experiment; he hopes people who grew up with these tools will find uses that established teams had not imagined.