How Intercom 2X'd engineering velocity with Claude Code | Brian Scanlan
Brian Scanlan from Intercom explains how they doubled their R&D team's pull request throughput in 9 months by implementing Claude Code across their engineering organization. He demonstrates their comprehensive approach including skills development, telemetry tracking, and quality controls that enabled 2x velocity gains while maintaining code quality.
Summary
Brian Scanlan, a senior principal engineer at Intercom, discusses how his company achieved a 2x increase in engineering throughput by fully embracing AI-assisted coding with Claude Code. Intercom made this transformation after recognizing AI as an existential opportunity following ChatGPT's release, with leadership setting ambitious goals for doubling R&D output measured by pull requests per R&D employee.
The implementation involved creating a comprehensive skills repository with hundreds of custom tools, from basic PR creation to complex tasks like fixing flaky tests. They built extensive telemetry using Honeycomb and session analysis to track skill usage and developer progress. Key innovations include automated PR description quality control, CI/CD integration, and treating their internal AI adoption like a product with proper instrumentation.
Scanlan demonstrates their approach by live-coding a simple redirect feature, showing how their system enforces quality standards through hooks and skills. He explains their philosophy of 'agent-first' work and removing barriers to AI adoption while maintaining high engineering standards. The session includes discussion of their flaky test fixing skill, which evolved from basic automation to a sophisticated tool that can handle complex debugging scenarios.
The conversation extends to customer-facing implications, with Scanlan showing how they're building CLI tools and agent-friendly interfaces for their products. He emphasizes the cultural transformation, describing increased excitement and productivity across teams, while acknowledging significant token costs that they treat as worthwhile investment. The approach has enabled engineers to tackle previously intractable technical debt and achieve 'backlog zero' on long-standing issues.
About this episode
Brian Scanlan is a senior principal engineer at Intercom, where he’s led the company’s transformation to AI-first engineering. In just nine months, Intercom doubled their R&D throughput while maintaining code quality, with 100% of engineers—plus designers, PMs, and TPMs—now shipping code via Claude Code. *What you’ll learn:* 1. How Intercom doubled their merged PRs per R&D employee in just nine months using Claude Code 2. The telemetry infrastructure they built to measure AI adoption and quality across hundreds of engineers 3. Why they built a skills repository with hooks that enforce engineering standards automatically 4. How they’re preparing their product for an agent-first world with CLIs, MCPs, and ephemeral APIs 5. The permission and accountability framework that enabled rapid AI adoption 6. Why backlog zero is now achievable and what that means for engineering culture *Brought to you by:* Celigo—Intelligent automation built for AI: https://celigo.com/howIAI Cursor—The best way to code with AI: https://www.chatprd.ai/howiai *In this episode, we cover:* (00:00) Introduction to Brian Scanlan (02:40) Why Intercom went all-in on AI for both product and engineering (05:01) The breakthrough moment with Opus 4.6 and Christmas break 2025 (07:02) Demo: Intercom’s merged PRs per R&D head (12:50) Agent-first work as a fundamental reimagining of technical workflows (14:27) The cost tradeoff: treating AI spend as an investment (16:47) Measuring quality (21:22) Demo: Shipping a redirect in the Rails monolith with Claude Code (24:03) Creating a custom PR skill (26:33) Building a software factory with predictable quality standards (30:15) Telemetry infrastructure: Honeycomb for skill usage tracking (32:10) Session data collection and personalized usage insights (36:08) Quick overview (39:20) Walking through Intercom’s skills repository (42:16) Deep dive: The flaky spec skill and how it reached 100x capability (46:44) The “and then” workflow for building comprehensive skills (52:31) The live website and overview of workflows (53:32) How internal AI experience informs customer product decisions (56:18) Making SaaS products agent-friendly with CLIs and helpful hints (01:03:49) Why conversion drop-off is invisible in agent-driven workflows (01:05:28) Lightning round and final thoughts *Detailed workflow walkthroughs from this episode:* • How Intercom Doubled Engineering Output: Brian Scanlan's 4 AI Workflows for Claude Code: https://www.chatprd.ai/how-i-ai/how-intercom-doubled-engineering-output-brian-scanlan-ai-workflows-for-claude-code • Design an Agent-Friendly CLI to Automate SaaS Product Onboarding: https://www.chatprd.ai/how-i-ai/workflows/design-an-agent-friendly-cli-to-automate-saas-product-onboarding • Build a Self-Improving AI Agent to Automatically Fix Flaky Tests: https://www.chatprd.ai/how-i-ai/workflows/build-a-self-improving-ai-agent-to-automatically-fix-flaky-tests • Automate High-Quality Pull Request Descriptions with a Custom AI Skill: https://www.chatprd.ai/how-i-ai/workflows/automate-high-quality-pull-request-descriptions-with-a-custom-ai-skill *Tools referenced:* • Claude Code: https://claude.ai/code • Cursor: https://cursor.com/ • Honeycomb: https://www.honeycomb.io/ • Snowflake: https://www.snowflake.com/ • Fin AI: https://www.intercom.com/fin • Vercel: https://vercel.com/ *Other references:* • Intercom GitHub Repo: https://github.com/intercom • Google API Go Client Repo: https://github.com/googleapis/google-api-go-client *Where to find Brian Scanlan:* X: https://x.com/brian_scanlan LinkedIn: https://www.linkedin.com/in/scanlanb/ Company: https://www.intercom.com *Where to find Claire Vo:* ChatPRD: https://www.chatprd.ai/ Website: https://clairevo.com/ LinkedIn: https://www.linkedin.com/in/clairevo/ X: https://x.com/clairevo _Production and marketing by https://penname.co/._ _For inquiries about sponsoring the podcast, email [email protected]._
Key Insights
- Intercom achieved a 2x increase in pull requests per R&D employee in 9 months after implementing Claude Code organization-wide
- Scanlan argues that imagination, not tools, became the primary barrier to productivity after AI models like Opus reached sufficient capability
- The company uses LLM judges to evaluate pull request description quality and found that AI-generated descriptions were initially terrible, requiring custom skills to maintain standards
- Intercom distributes AI skills through internal IT systems rather than Claude Code's plugin mechanism because they found the official system too flaky for enterprise deployment
- Scanlan claims their flaky test fixing skill evolved from basic automation to distinguished engineer-level capability by incorporating self-learning and progressive discovery
- The team treats AI token costs like hiring whole new offices of people but considers it worthwhile investment during the current alpha phase
- Scanlan reports that Stanford research group analysis of their code showed quality was actually improving with increased AI usage, contrary to common concerns
- He argues that 'backlog zero' is now realistic for engineering teams because the cost and time barriers to tackling technical debt have been dramatically reduced
- Scanlan describes re-implementing a Go microservice in Ruby as a single Claude Code session, work that previously would have required roadmap planning and team coordination
- The company believes all technical work will become 'agent first' and is setting deadlines for this transition rather than gradual adoption
- Scanlan argues that SAS products need to become more agent-friendly with better CLIs, APIs, and self-service capabilities to remain competitive
- He reports having 'the most amount of fun in my career over the last 3 months' due to the ability to quickly realize previously impossible ideas
Topics
Transcript
[0:00] Suddenly you started realizing that you have to think bigger about things or that your imagination is now the barrier not the tool. >> How is this not happening in your organization? Like literally the physical limits of my ability to type code are unlocked by AI. >> Today we are seeing twice the number of throughput as we did compared to 9 months ago on our engineering team. Now it's like why can't it be 10x? This is a little bit more of what my instinct tells me is possible, which is if you go allin, if you prepare your team, if you prepare your codebase, I think your overall product quality is going to go [0:31] up.…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
GPT-5.6's video editing via Codex is genuinely one of my favorite new workflows
A product manager describes using GPT-5.6 with Codex to automate video editing for social media clips. By simply dragging a file and providing natural language instructions, the AI generated five polished, fast-paced hype videos from a long conference talk recording, dramatically reducing the time-intensive manual clipping process.
Theoretically Intelligent vs. Practically Effective: Why GPT-5.6 Sol Beats Fable for Product Work
An executive contrasts two AI models (Fable and Soul/GPT-5.6 Sol), arguing that Soul is superior for product work because it prioritizes practical effectiveness over theoretical intelligence. The speaker values the ability to ship products to customers and understand end-user goals over theoretical sophistication.
Build a harness when the same workflow needs the same setup and the same outcomes, every time
Building a harness for repetitive workflows allows you to be more prescriptive about job execution, resulting in greater efficiency, consistency, and better outcomes. Rather than explaining requirements to an AI agent each time, a harness lets you use a simpler interface like pasting a link while the agent already understands the intended task.
GPT 5.6-Sol vs. Claude Fable: Why OpenAI’s new model crushes my benchmark
Claire Vo compares OpenAI's new GPT 5.6 models (Soul, Terra, Luna) against Claude's Fable using her custom "How I AI" benchmark, finding that GPT 5.6 Soul excels at practical product work, prototyping, and natural communication, while Fable is theoretically intelligent but pedantic and difficult to collaborate with.
Context offloading is an underrated AI use case
The speaker highlights context offloading as an underrated AI use case, where AI serves as a safety net for routine cognitive tasks like email management and personal finances. Rather than adding new capabilities, AI reduces anxiety about missing important information or making mistakes by handling monitoring tasks, thereby freeing up mental bandwidth.