Intercom and Claude Code: A Revolution in Software Engineering

Brian Scanlan, Principal Engineer at Intercom, shared how the company managed to double its engineering velocity in just nine months through the integration of Claude Code. This transformation was made possible by adopting custom skills, in-depth telemetry, and a culture of permission, turning AI into a true force multiplier.

A Product Approach to Engineering

Intercom has adopted an approach where its engineering organization is treated like a product. Every aspect is instrumented, using Honeycomb to track skill invocations and storing anonymized Claude Code sessions in S3. Custom dashboards have been created, allowing engineers to compare themselves with their peers. This approach is not surveillance but an application of product thinking to customer-oriented features. Visibility into what works and what doesn’t is crucial for improving and expanding AI adoption.

Strong Preparation and Infrastructure

The 2x velocity gain was only possible with a solid infrastructure in place. Intercom doubled the number of merged PRs per R&D employee in nine months, thanks to a mature CI/CD process, comprehensive test coverage, and a high-trust culture. AI amplifies existing strengths but could have accelerated the shipping of faulty code without a solid foundation. If the deployment pipeline is faulty or if the code review process is chaotic, AI will only speed up these issues.

Custom Skills and Code Quality

Custom skills with hooks ensure quality at the point of creation. For example, the "Create PR" skill prevents direct use of the GitHub CLI, forcing Claude Code to write context-rich PR descriptions. This approach has improved code quality while increasing delivery speed. Intercom's partnership with researchers from Stanford has shown that their code quality metrics are increasing, not decreasing.

Culture of Permission and Innovation

The role of technical leadership has been to grant permission to experiment. Brian Scanlan has a simple framework: tell people they can do things, and if something goes wrong, blame me. Engineers don’t need more tutorials or documentation; they need permission to connect Claude Code to tools like Snowflake, to ship code from their phones on the subway, or to build a CLI that bypasses email verification. This approach has allowed Intercom to effectively manage technical debt and improve the developer experience. The cost of fixing unstable tests, enhancing the developer experience, and managing technical debt is approaching zero.

Towards an Agent-Oriented Future

Intercom has also developed a CLI that autonomously registers at Fin, verifies email addresses, and completes installations without human intervention. Brian Scanlan envisions a future where all work becomes agent-oriented, and he recommends setting a deadline for this. His vision is that by the end of each month, the first response to an alert, a planning meeting, or a customer question should be an agent performing the basic work. This is not an aspiration but a realistic expectation given the current state of models and infrastructures.

In conclusion, Intercom's experience with Claude Code illustrates how thoughtful integration of AI can transform engineering practices, multiplying both the efficiency and quality of software development. This approach demonstrates that when the cost of fixing unstable tests, enhancing the developer experience, and managing technical debt approaches zero, it becomes possible to truly tackle these issues rather than just discussing them during retrospectives. The business constraint on internal projects disappears when agents can execute them in hours instead of quarters, thereby transforming traditional workflows and freeing teams to focus on innovation.