Coinbase AI Adoption: Ditching Meetings for PR Velocity

The AI Acceleration Playbook: How Coinbase Ditched Meetings for Massive PR Velocity

In a world where AI adoption in large engineering organizations is met with skepticism, Chintan Turakhia, Senior Director of Engineering at Coinbase, offers a compelling counter-narrative. This conversation reveals that the true power of AI lies not just in its ability to generate code, but in its capacity to dismantle deeply ingrained inefficiencies and coordination overhead. The hidden consequence of embracing AI, as Turakhia demonstrates, is a radical shift in leadership from managing by meetings to enabling by doing, freeing up significant engineering time for actual product development. This analysis is crucial for engineering leaders, VPs of Engineering, and CTOs who are grappling with how to translate AI hype into tangible velocity gains, offering them a blueprint to unlock competitive advantage through accelerated feedback loops and empowered engineers.

The "Adopt or Die" Imperative: Shifting Engineering Culture with AI

The sheer scale of Coinbase's engineering organization--over a thousand engineers--makes the success of their AI adoption strategy particularly noteworthy. Turakhia’s approach isn't about philosophical debates on AI's future; it's about immediate, hands-on application to solve pressing business needs. Faced with the daunting task of rewriting a core product into a consumer-facing social app within an aggressive six-to-nine-month timeline, the team turned to AI not as a luxury, but as a necessity. This wasn't about incremental improvements; it was about a fundamental acceleration of velocity to compete with established giants. The initial skepticism, mirrored by many organizations, stemmed from early, less capable AI tools. However, Turakhia’s conviction, coupled with a deep dive into tools like Cursor, became the catalyst for change.

"I think it's not only possible, it's adopt or die."

This statement underscores the urgency. The conventional wisdom of gradual tool adoption or top-down mandates proved ineffective. Instead, Turakhia championed a hands-on, "show, don't tell" leadership style. By personally immersing himself in AI tools, he not only rediscovered the joy of coding but also identified practical, albeit sometimes mundane, use cases--like summarizing interview notes or drafting PR descriptions--that directly addressed engineer "shit work." This focus on reducing toil, from linting to unit test generation, became the gateway to broader adoption. The creation of a dedicated "Cursor Wins" Slack channel fostered a community of practice, where early successes were shared, building momentum organically.

The "PR Speed Run": A Shockwave of Velocity

The most striking manifestation of this AI-driven transformation was the "PR speed run." This wasn't just a team-building exercise; it was a deliberate, timed event designed to break through inertia and demonstrate immediate, tangible output. By setting a clear goal--to push as many PRs as possible in a short, focused period--the team achieved remarkable results, generating 70 PRs from 100 engineers in just 15 minutes in one instance, and later, 300-400 PRs from 800 engineers in 30 minutes during a company-wide event. This not only pressured their infrastructure (breaking GitHub, which Turakhia saw as a positive stress test) but also fundamentally altered the team's perception of what was possible.

"This is a moment where we should be breaking the rules because AI is breaking the rules for us."

This quote encapsulates the paradigm shift. When AI fundamentally changes the cost and effort of certain tasks, the old rules of engagement--the lengthy approval processes, the fear of shipping minor changes--become obsolete. The speed run demonstrated that by leveraging AI, teams could bypass traditional bottlenecks, leading to a dramatic increase in merged PRs and, crucially, a compressed cycle from initial idea to customer feedback. This wasn't about replacing engineers; it was about augmenting them, freeing them from drudgery to focus on higher-value work and innovation.

From Toil to Trade: AI-Powered Feedback Loops

The transformation of the feedback-to-feature cycle is another critical consequence-mapping insight. Traditionally, user feedback collection was a multi-step, often slow process involving manual transcription, triage, and prioritization. Turakhia's team built an in-house tool that captures unstructured audio feedback from user sessions, transcribes it, identifies bugs and suggested fixes, and automatically generates Linear tickets. This process, which previously could take days or weeks, now happens in minutes.

"The cost of writing something in Slack is zero, but the cost of answering something in Slack is enormous, and most of it is noise."

This observation highlights the core problem AI can solve: reducing the noise and friction in communication and workflow. By automating the initial stages of feedback processing, the team drastically cuts down the time to action. The subsequent step, creating a PR from the ticket, is also automated via a custom bot, Cloudbot. This bot, built in-house due to specific security requirements, integrates with tools like Linear and leverages underlying AI models to draft PRs, generate buildable code, and even provide deep links to Cursor for further refinement. This creates an almost instantaneous loop from customer insight to deployed code, a competitive advantage that conventional methods simply cannot match. The key here is not just automation, but the speed of automation, enabling a rapid iteration cycle that keeps pace with market demands.

Analyzing the AI User: Data-Driven Adoption

Beyond the direct application of AI in development, Turakhia’s team also used AI to analyze AI adoption itself. By feeding Cursor's analytics data into Cursor's own AI capabilities, they could identify distinct user cohorts--from "inactive" to "super users." This analysis moved beyond vanity metrics like "lines of code generated by AI" to understand how engineers were using the tools, identifying patterns of agent versus tab completion usage, and pinpointing behaviors of power users.

This data-driven approach allows for targeted interventions. Instead of generic training, guidance can be tailored to specific cohorts. For instance, inactive users might receive simple prompts to try basic features, while power users are encouraged to explore more advanced capabilities. This nuanced understanding of user behavior enables a more effective and efficient strategy for driving broader AI adoption, ensuring that the tools are not just available but actively and effectively utilized across the organization. The output of this analysis--playbooks, Slack posts, and visualizations--further democratizes AI knowledge and reinforces the cultural shift.

Key Action Items:

• Immediate Action (Next 1-2 Weeks):

  • Establish a "Wins & Losses" Channel: Create a dedicated Slack channel for engineers to share their experiences with AI tools, both successes and failures. This fosters transparency and community learning.
  • Identify "Toil" Tasks: Have teams catalog the most repetitive, soul-crushing tasks they currently perform (e.g., writing boilerplate tests, formatting code, basic documentation).
  • Pilot AI for Toil Reduction: Select 1-2 high-impact toil tasks and experiment with AI tools (like GitHub Copilot, Cursor, or custom agents) to automate them for a small pilot group.
  • Leadership Hands-On: Engineering leaders should dedicate at least 1-2 hours per week to personally using AI tools for coding or workflow tasks, documenting their learnings and sharing them.

• Short-Term Investment (Next 1-3 Months):

  • Implement "PR Speed Run" Events: Organize short, timed events where engineers are encouraged to pick up small, neglected tasks and push PRs using AI assistance. Aim for 1-2 such events per quarter.
  • Explore AI-Assisted Feedback Capture: Investigate tools or build simple in-house solutions to automate the transcription and initial triage of user feedback (audio, video, or text).
  • Analyze AI Adoption Data: If using tools like Cursor, leverage their analytics or build simple scripts to understand usage patterns and identify power users and adoption bottlenecks.

• Longer-Term Investment (6-18 Months):

  • Develop Custom AI Agents/Workflows: For critical, repetitive workflows (e.g., automated PR generation from tickets, custom Slack bots for task automation), explore building or integrating custom AI agents. This requires dedicated engineering effort but yields significant downstream benefits.
  • Integrate AI into Core Feedback Loops: Fully automate the feedback-to-feature cycle, from capture and triage to PR generation, creating a rapid iteration loop that delights users and outpaces competitors.
  • Foster a "Super Builder" Culture: Identify and empower individuals or small teams to become internal champions for AI tooling, tasked with creating more AI-powered workflows and fostering broader organizational adoption. This pays off by creating a self-sustaining ecosystem of AI-driven productivity.