Intercom doubled merged PRs per R&D employee in 9 months. Senior principal engineer Brian Scanlan architected the shift, putting Claude Code in the hands of every engineer, designer, PM, and TPM at the company. 100% of those roles now ship code. The catalyst was a specific inflection point: Opus 4.6 over Christmas break 2025, when Scanlan says the model's capability crossed a threshold that made the organizational bet obvious.
The implementation details are what make this worth reading in full. Intercom built custom telemetry on Honeycomb to track skill usage across hundreds of engineers, constructed a skills repository with hooks that programmatically enforce engineering standards, and developed a permission and accountability framework that let adoption move fast without quality collapse. They treat AI spend as a capital investment, not an operating cost. The flaky spec skill alone reached what Scanlan calls 100x capability through an iterative 'and then' workflow for building comprehensive agent instructions.
The forward-looking section hits harder than most. Intercom is rebuilding its own product for an agent-first world: CLIs, MCPs, and ephemeral APIs are the target interface layer. Scanlan also argues that backlog zero is now a realistic engineering state, which has direct implications for how teams are structured and how work gets prioritized. The full conversation includes live demos of the Rails monolith workflow and a walkthrough of the skills repository.
[READ ORIGINAL →]