CHAPTERS
- 0:10 – 1:10
Why DoorDash rolled out Cowork + Claude Code company-wide
Andy Fang describes a major rollout aimed at raising the baseline of AI fluency across DoorDash, not just among engineers. He explains how many employees still think of AI as “just chat,” and how connecting AI to everyday tools (email, calendar, Slack) can unlock big productivity gains.
- •Goal: raise the floor of AI fluency across the whole company
- •Non-engineers and even execs were underutilizing AI capabilities
- •Integrations with Gmail/Calendar/Slack make AI useful for knowledge work
- •Early impact: measurable increases in organizational throughput
- 1:10 – 1:58
Andy’s coding origin story—and his return to shipping code
Andy recounts learning to code at age nine and writing DoorDash’s early code himself in the Stanford dorm days. As the company scaled, he stopped coding—until Claude Code enabled him to ship production code again.
- •First coding exposure at a childhood coding camp
- •Early DoorDash: Andy wrote production code directly
- •Scaling shifted his role away from hands-on coding
- •Claude Code triggered a “comeback” to shipping production code
- 1:58 – 2:49
Hands-on workflow: terminal-first, ‘no manual code,’ and parallel sessions
Andy explains his personal Claude Code setup: primarily in terminal, sometimes desktop, with an explicit goal of not writing code manually. He also describes keeping multiple sessions and repos active using worktrees to avoid conflicts.
- •Uses Claude Code primarily via terminal (sometimes desktop app)
- •Personal rule: avoid writing code manually—agent writes everything
- •Runs multiple repos and multiple sessions concurrently
- •Uses worktrees to prevent conflicts when parallelizing work
- 2:49 – 3:41
Model inflection point: from struggling with setup to shipping in 5 languages
Andy contrasts an early attempt (couldn’t even configure the local environment) with a later attempt where “it just worked.” He attributes this to a capability inflection point in newer models, enabling real production delivery across multiple languages.
- •Initial attempt failed due to environment/config complexity
- •Re-tried later (early 2026) and it worked end-to-end
- •New models are dramatically better at ‘figuring things out’
- •Andy reached shipping production code across five languages
- 3:41 – 5:40
Unlearning attachment: throwing away old approaches as models improve
Boris and Andy discuss a new pattern: retrying the same idea months later can succeed because models evolve. Andy emphasizes that teams must be willing to discard previously “working” setups and redesign workflows around fast-changing capabilities.
- •Old-model failures can become new-model successes
- •Teams should be willing to throw away prior tooling/workflows
- •Leaders should personally ship to understand real constraints
- •Focus on production delivery, not just local prototypes
- 5:40 – 6:51
New bottlenecks: merge queues, CI/CD, code review, and security
With higher code output, Andy notes that the limiting factors shift to merging and operational processes. DoorDash is investing in AI code review and exploring AI-assisted security checks as security issues rise industry-wide.
- •AI increases output but creates merge/review bottlenecks
- •CI/CD and release processes need reimagining for new velocity
- •AI code review agents help scale review capacity
- •Security issues are increasing; AI must also help catch them earlier
- 6:51 – 8:09
Adoption timeline and governance: fast access with strong security guardrails
Andy explains DoorDash introduced Claude Code in 2025, with an inflection point in adoption late 2025/early 2026. He highlights tight partnership across executive leadership, IT, security, and engineering to provide rapid access while keeping security non-negotiable.
- •Introduced Claude Code in 2025; adoption surged late 2025/early 2026
- •Created procurement/review processes tailored for fast-moving AI tools
- •Optimized early for exploration over strict budget limits
- •Security partnership and guardrails were central from the start
- 8:09 – 9:31
Creating psychological safety: champions, shared wins, and sharing failures
Andy describes how DoorDash encouraged experimentation by letting early adopters lead, then amplifying success stories. He stresses the importance of written artifacts and also normalizing the sharing of failed workflows and wasteful experiments.
- •Start by distributing tools widely and watching for early adopters
- •Turn success cases into internal advocates within teams
- •Encourage written artifacts to scale learnings across the company
- •Share failures too (e.g., token-wasting integrations) to accelerate learning
- 9:31 – 10:58
How DoorDash uses Claude Code today: org-wide usage and the Flux platform
Claude is now used broadly either through Claude Code or via Claude models embedded in other tools. Andy introduces Flux, DoorDash’s internal platform of security-approved cloud VMs for running Claude sessions, powering automation like AI code review agents.
- •Widespread Claude usage across the organization
- •Investment focus: automation to capitalize on increased throughput
- •Flux: internal, security-blessed cloud VM platform for Claude sessions
- •Built AI code review agents using Agent SDK + Claude models
- 10:58 – 12:37
Cross-functional acceleration: designers shipping code and the 3–5× challenge
DoorDash pushed designers to ship production code to embed more directly into the development cycle. Andy describes setting an org challenge: complete projects 3–5× faster, then studying which constraints (technical and organizational) prevent achieving that speed.
- •Designers were encouraged to ship to production (learning-by-doing)
- •Best teams embed design/PM tightly into the dev cycle using Claude
- •Org-wide goal: finish projects in 3–5× less time with AI
- •Collect lessons from high-performing teams and publish playbooks
- 12:37 – 15:29
What works vs. what breaks: small teams, agent-friendly repos, and process friction
Andy shares concrete adoption learnings: smaller teams accelerate, and making a codebase ‘agent-friendly’ pays off. The biggest friction appears in user-facing work where cross-functional reviews (product/design/ship) become the new critical path.
- •Smaller teams reduce coordination overhead and move faster
- •Agent-friendly repos: codify architecture principles in repo markdown
- •Standardized ‘skills’ (e.g., mobile simulator/test workflows) boost leverage
- •User-facing features hit cross-functional alignment/review bottlenecks
- 15:29 – 16:57
Rethinking org design: self-sufficiency, fewer gatekeepers, more generalists
Andy argues AI velocity requires redesigning how teams collaborate and who owns what. He suggests fewer specialist gatekeepers, more generalist mobility across the codebase, and using agents plus documented principles to enforce quality constraints.
- •AI makes full-stack mobility more achievable for specialized roles (e.g., mobile)
- •Keep a smaller set of true domain experts; reduce gatekeeping burden
- •Codify expert knowledge as principles/automated checks agents can use
- •Promote self-sufficient individuals/teams to unlock speed
- 16:57 – 19:06
Staffing and sponsorship: giving teams permission to break constraints
To enable real experimentation, Andy emphasizes executive sponsorship and explicit permission to surface non-coding blockers. Without changing the surrounding constraints, teams tend to abandon AI-first rework quickly.
- •Create ‘safe’ teams with VP-level sponsorship to unblock quickly
- •Don’t expect breakthroughs if teams must operate under old constraints
- •Give token budget plus mandate to identify cross-functional blockers
- •Scale via sharing artifacts and letting practices spread organically
- 19:06 – 22:39
Measuring ROI: throughput vs. customer value, and the frontier of knowledge work
Andy discusses how ROI is clearer in engineering (throughput, delivery speed, fewer people per project) but still emerging in knowledge work. He frames two knowledge-work angles: raising baseline productivity across everyone and automating high-leverage workflows by domain.
- •Engineering ROI: code/delivery speed (imperfect metrics) and staffing leverage
- •Ultimate metric: faster customer value delivery, not just merged code
- •Knowledge work ROI: baseline time savings across common tasks
- •Also target domain workflows (e.g., sales QBRs) for bigger step-function gains
- 22:39 – 25:11
Finding ‘AI champions’ and closing advice for leaders and new grads
Andy explains DoorDash identifies champions by giving tools broadly and observing who becomes prolific and vocal. He closes with advice: leaders should use AI themselves, support identity/workflow change, and rely on written artifacts; new grads should stay curious and experiment to discover new paradigms.
- •Champions emerge organically when tools are widely available
- •Champions help uncover domain pain points leadership may not know
- •Advice to leaders: use AI personally; build empathy for workflow/identity shifts
- •Advice to leaders: written artifacts scale learning and feed agents
- •Advice to new grads: stay curious and push new AI-native ways of working
