How Building with AI Can Double the Throughput of Your Engineering Team — Brian Scanlan, Intercom
Intercom doubled its engineering throughput in under a year by treating its AI coding agent not as a simple tool, but as a new senior engineer. This involved a full onboarding process onto their 15-year-old Rails monolith, creating a library of durable skills for recurring tasks, and providing audited access to all internal systems.
Intercom has fundamentally redefined its software development lifecycle by treating its AI coding agent, Claude, not as a tool for simple tasks, but as a new senior engineer joining the team. This strategic shift, initiated in mid-2023, led to a doubling of engineering throughput in less than a year, moving beyond the marginal gains seen from initial explorations with tools like GitHub Copilot and Cursor.
The “2x” Initiative: A Mandate for Change
The project, internally named “2x,” was driven by a strong conviction that AI would transform all knowledge work. The primary goal was to double engineering throughput, measured by code changes per R&D person. This ambitious target was backed by decisive executive guidance and a comprehensive organizational change strategy:
- A Full-Time Team: A dedicated, and growing, team of top engineers was tasked with leading the AI integration effort.
- Clear Expectations: Job descriptions were updated across R&D. Failing to adopt AI became synonymous with not meeting expectations.
- Cultural Reinforcement: The company actively celebrated successes in public Slack channels, ran hackathons, and held AI immersion days to bring the entire R&D organization along on the journey.
The Platform Strategy: All-In on a Single Agent
A critical decision was to standardize on a single platform, Claude Code, for all engineers. This approach was chosen to avoid “model anxiety” and the fragmented, non-compounding benefits of a multi-tool environment. The vision was to create a deeply integrated system where the agent could function as a senior engineer on any technical task at Intercom.
Onboarding an AI Agent as a Senior Engineer
The core of Intercom’s success lies in its methodology for integrating the agent, which mirrors the process of onboarding a human engineer.
- Connecting to Everything: The agent was given access to the same systems as a human engineer, including production environments. This was done with confidence, leveraging Intercom’s mature controls, permissions, and audit trails. The guiding principle is: “Anything I do on my laptop, Claude should be able to do.”
- Teaching the “Intercom Way”: The agent was systematically taught the company’s specific conventions, architectural patterns, and best practices developed over 15 years, including its Ruby on Rails monolith, React patterns, testing standards, and security rules. This knowledge is encapsulated in a library of “engineering captures,” “skills,” and “guidance.”
- Developing Durable Skills: Instead of building complex, opinionated orchestrators, the focus is on creating small, high-quality, durable, and testable skills that perform specific jobs extremely well. When the agent fails or goes down the wrong path, the team updates the guidance, creating a continuous improvement flywheel that everyone contributes to.
Guiding Principles for Human-Agent Collaboration
To steer this new way of working, Intercom established several key principles:
- Give Agents Problems, Not Tasks: Engineers are encouraged to describe the problem they are facing and let the agent determine the correct sequence of skills to invoke. Brian Scanlan shared an anecdote where he was pulled into a security incident involving a potential data leak. By simply describing the situation to Claude, the agent autonomously invoked an internal skill he didn’t know existed, downloaded the relevant files, performed a full analysis based on company policy, concluded the data was innocuous, and provided the next steps—all within two minutes.
- Moving Up the Stack: The role of the engineer is evolving. Much like the transition from a sysadmin racking servers to an SRE focused on automation, engineers are moving away from direct code implementation and toward a role of guiding, automating, and optimizing the environment for AI agents.
Quantifiable Results and Impact
The “2x” initiative has produced significant, measurable results:
- 2x PR Throughput: The primary goal was achieved in under a year, with a dramatic inflection point in pull request volume seen immediately after the company standardized on a single tool in January.
- Automated Code Reviews: The increased PR volume created a new bottleneck in code review. To address this, Intercom built an automated approval system. Currently, 17.6% of pull requests are automatically approved. This system is fully compliant with SOC 2, ISO 27001, and HIPAA, having been rigorously backtested against historical data and proven to reduce risk compared to human review for certain classes of changes.
- Systemic Improvements: The company has seen a natural decline in open defects, as they are now closed faster than ever. Furthermore, code quality, as measured by a Stanford research group, has been steadily increasing.
- Infrastructure Strain: The massive increase in development velocity “melted” the CI infrastructure, requiring a significant effort to scale it to handle the new load.
The entire process is supported by a robust data feedback loop. All session transcripts are ingested into S3 for data mining, which helps the team understand skill usage, identify effective patterns, and continuously improve the agent’s performance. This approach represents a paradigm shift from using AI as a simple code completion tool to integrating it as a core, productive member of the engineering team.