How Coinbase scaled AI to 1,000+ engineers | Chintan Turakhia

High-level takeaway

Key idea: AI is an accelerant (not a headcount replacement). Success required hands-on leadership, showing (not just mandating) how to use tools, removing toil, measuring output, and surfacing quick wins.

• Coinbase scaled AI tooling and workflows across 1,000+ engineers to dramatically increase velocity, reduce coordination overhead, and make AI an ingrained engineering productivity multiplier — not a philosophical experiment.
• The emphasis was on practical, measurable improvements (time-to-action, review time, merge → deploy) rather than abstract goals.

What they rebuilt and why

• Rewrote a product (Coinbase Wallet → a consumer social crypto app) on a tight 6–9 month timeline with a smaller, high-performing team using React Native.
• Needed huge velocity and pragmatic tooling to compete with much larger incumbents.

Organizational / change tactics that worked

• Single high-conviction leader who’s hands-on in the code to model usage and iterate.
• “Show, don’t tell”: leaders demoed AI workflows and shipped code themselves to lower the barrier for others.
• Focus on removing repetitive, soul-sucking engineering toil (linting, unit tests, PR descriptions, small bug fixes).
• Social proof and low-friction sharing: channels like “cursor-wins” (and wins+losses) to surface examples and troubleshooting.
• Speedruns / surges: time-boxed events where engineers fix trivial items en masse to produce immediate, visible wins (examples: 100 people → ~70 PRs in 15 minutes; companywide ~800 engineers → 300–400 PRs in 30 minutes). These events also stress-test tools/infrastructure and boost morale.
• Encourage shipping and break rules when necessary — AI enables quicker, lower-cost experiments.

Metrics and outcomes

• Primary success metric: time from ticket → change in production (time-to-action, review time, merge → deploy).
• Example outcome: PR review cycle time dropped ~10x (from ~150 hours to ~15 hours) after introducing AI-assisted review flows and other changes.
• Use cohort analytics (light → power/super users) to identify and replicate power-user behaviors.

Tools, integrations and technical patterns

• Cursor (used heavily): agent mode, tab completions, cursor rules, Bugbot, analytics export (CSV).
  • Use cursor rules to automate repetitive tasks (create draft PRs, PR descriptions, unit tests).
  • Use cursor analytics to identify cohorts and generate playbooks.
• In-house agents + Cloudbot: built internal agent (Cloudbot) that:
  • Captures live user feedback (audio/video), runs a small LLM to summarize/triage,
  • Creates Linear tickets automatically,
  • Creates PRs/branches and returns deep links/QR codes for QA.
• Data/context pipeline: Linear as the canonical context store; agents pull additional signals from DataDog, Sentry, Amplitude, Snowflake, etc. Context is critical for useful agent actions.
• Deep integration with Slack to make workflows visible and viral (Slack → Linear → Cloudbot → PR).
• When security/compliance restricts cloud agents, building internal agents is realistic and viable.

Concrete workflows and automation examples (playbook-style)

• PR speedrun
  1. Everyone picks trivial tickets.
  2. Run AI-assisted workflow to create draft PRs.
  3. Assign reviewers/testers and merge fixes rapidly.
• Live feedback capture
  • Record audio/video in a “feedback cafe” → LLM extracts bugs → create Linear tickets → agent creates PR → QA via QR link.
• Cohort analysis pipeline
  • Export Cursor CSV → ask an LLM to cluster users into cohorts (inactive/light/regular/power/super) → generate a Python script and HTML dashboard → produce a Slack playbook + suggested prompts to move users between cohorts.
• Playbooks & gamification
  • Slack posts, short playbooks, “quests” for new users, and leaderboard-like incentives to nudge adoption.

Practical career / hiring notes

• Create a “super builder” role: hire people whose job is to create tooling/agents that make other builders more productive. This offers clear career upside for engineers who want to lead AI adoption.
• Advice for engineers: learn to build/integrate agents and pipelines; be an early AI advocate in your organization.

Security / enterprise considerations

• Agent/copilot tools for enterprise apps must satisfy auth, access controls, and audit logs.
• Work OS–style platforms can provide enterprise feature APIs to reduce implementation burden.
• If compliance prevents using third-party agents, custom in-house agents + internal tooling + secure data pipelines are a valid path.

Examples of consumer / personal use cases shown

• Capturing school event emails → auto-generate calendar invites.
• Wine/champagne taste profiling: feed tasting notes/images → LLM extracts taste profile and recommends bottles from a menu (demonstrates personalization and reverse-engineering preferences).

Lessons learned / cultural impacts

• Leaders regained coding time and cut meeting overhead by trusting teams to act (lighter calendars).
• Rapid feedback loops increase fun and iteration speed; the dopamine from frequent shipping accelerates adoption.
• Always center context; shallow prompts without context give poor results.

Guides / repeatable recipes

• Run short, visible speedrun events to create conviction and viral adoption.
• Automate low-value tasks first (lint, tests, PR prose) to demonstrate value quickly.
• Use Cursor or similar tools for quick automation plus analytics; export usage data and run cohort analysis with LLMs to generate playbooks.
• Build a Slack + ticketing + agent loop (capture → summarize → ticket → PR) for immediate feedback-to-feature cycles.
• Create “wins+losses” channels and public playbooks to share what worked and what didn’t.

Sponsors / tools called out

• Cursor (agent mode, Bugbot, analytics)
• Linear (ticketing; good agent integrations)
• GitHub (used for PRs; noted to have broken under load)
• Work OS (sponsor discussed: drop-in enterprise features)
• Atlassian / Rovo (ad for AI teammate built into Jira/Confluence)
• Internal: Cloudbot (Coinbase in-house agent)

Main speakers / sources

• Claire Bell — host (product leader, “How I AI”).
• Chintan Turakhia — Senior Director of Engineering at Coinbase (primary guest and source of the practical examples and tactics).

(Note: subtitles were auto-generated; some names/words may be slightly off in transcription.)