Intercom 2x's Eng Throughput by Onboarding Claude Code

The Breakthrough

Intercom doubled engineering throughput by onboarding Claude Code to their 15-year-old Ruby on Rails monolith, writing durable skills for recurring tasks like fixing flaky specs, connecting it to production systems and internal tools, and requiring all engineers to use one platform instead of scattered tools like Cursor or Augment.

What Actually Worked

• Intercom staff dedicated a full-time 2x team to roll out Claude Code, updated job descriptions to make AI adoption mandatory ("not meeting expectations" without it), ran hackathons and AI immersion days, and celebrated successes in Slack channels to drive adoption across hundreds of engineers.
• They onboarded Claude Code like a new hire by encapsulating Rails conventions, architecture patterns, React standards, testing rules, and security policies into context, skills, guidance, and hooks; they pushed internal plugins to all laptops and connected it to production environments with controls, permissions, and audits.
• Engineers give Claude Code problems, not tasks; for example, during a security incident with accidentally published Snowflake table metadata in a public GitHub repo, Brian described the situation, joined a Slack channel, and Claude automatically invoked an unknown skill to download files, analyze against data breach policies, conclude it was innocuous, and provide next steps in two minutes.
• They built small, high-quality, testable skills using backtesting on historical code changes, incidents, and data; one skill fixes flaky specs across hundreds of thousands of tests using cheat codes, lookup tables, and progressive disclosure, performing at a level that impresses senior Rails engineers.
• Claude Code handles all technical work including debugging, testing, planning, and code review; they shaped PRs to be safe and simple for 17.6% auto-approval rate with SOC 2, ISO 27001, and HIPAA compliance via detailed backtesting, human labeling, and multimodal reviews.

Before / After

Pull request throughput doubled in under a year after mandating Claude Code in January. Auto Claude Code pull requests reached the 90s percentile. 17.6% of pull requests received automatic approval. Defects increased initially but now close faster than ever, with some teams achieving backlog zero. Code quality increased per Stanford research group metrics. CI infrastructure collapsed under volume.

Context

Intercom, a 15-year-old B2B SaaS company pivoting to AI with products like Fin (AI agent resolving 2M conversations weekly using custom models outperforming Sonnet), obsessed over developer productivity via metrics like code changes per R&D person. Dissatisfied with scattered tools like GitHub Copilot, Cursor, and Augment yielding only marginal gains, they launched the 2x project mid-last year amid model improvements, betting AI would transform knowledge work. Organizational changes, a single-platform bet on Claude Code, and a flywheel of continuous skill updates enabled agents to handle vast SDLC work, moving humans up the stack like the cloud did for sysadmins.

This matters because tools today suffice for agent-first workflows even without further model advances; Intercom's data shows risk reduction via well-defined agents outperforming inconsistent humans, with viral spread beyond engineering to product and design.

Notable Quotes

• "Give agents problems, not tasks."
• "All of engineering is changing: everything that you can do the agent must be able to do."
• "Our job is moving up the stack as engineers."
• "We're technically conservative: we like using single tools and just using them extremely well."
• "If you're not doing pretty much all of this today you're going to be doing it in the very near future."