Intercom doubled its engineering pull request throughput in less than a year by adopting Claude Code across the company and building a shared platform around it.
2
The team treated Claude Code like a new senior engineer, teaching it Intercom's Rails conventions, architecture, security rules, tools, and recurring engineering practices.
3
The largest gains came from giving agents broad problems to solve, then turning successful workflows into tested internal skills that other engineers could reuse.
Summary
Brian Scanlan describes Intercom's attempt to double engineering throughput without doubling its team. The company chose Claude Code as a shared platform, onboarded it to a 15-year-old Ruby on Rails monolith, connected it to internal systems, and wrote skills for recurring work. Scanlan argues that teams should give agents problems rather than narrowly defined tasks, allowing them to find the right skills and decide how to proceed. Intercom reports that pull request throughput doubled in less than a year. It also reports that 17.6% of pull requests are automatically approved under controls reviewed for SOC 2, ISO 27001, and HIPAA compliance. The approach created new constraints, especially in CI and code review, and required dedicated staff, executive direction, training, and constant iteration. Scanlan is candid that the work includes operational problems, uneven adoption, and a need to keep improving the skills and platform.
Intercom set an aggressive throughput target and measured code changes per R&D person
In the middle of last year, Intercom set out to double engineering throughput without doubling the team. Scanlan says the company measures many aspects of work through developer surveys and tools such as DX, but selected code changes per R&D person as its main productivity measure. He acknowledges that every measure has weaknesses once people optimize for it. The team still expected a large increase if AI was changing how work happened across the software development life cycle. The project was called 2x, and Scanlan describes the target as both wildly ambitious and reasonable given the direction of model and coding-harness capabilities.
AI adoption required executive direction, dedicated staff, and repeated organizational change
Intercom made AI adoption an explicit expectation for engineers, designers, and product managers. Scanlan says the company repeated the message across forums, celebrated useful automation in Slack, ran hackathons and AI immersion days, and staffed the effort full-time. The 2x team continued to grow because hundreds of people in R&D needed practical help rather than a general instruction to use AI. His advice for medium and large organizations is to assign strong people to this work full-time. The goal was to bring the organization along while changing everyday engineering practice.
Intercom chose one agent platform so its improvements would compound
After using GitHub Copilot, Cursor, Augment, and other tools, Intercom decided to standardize on Claude Code. Scanlan compares this to choosing a cloud platform: using one platform lets the company optimize it, improve it, and gain compounding benefits instead of spreading work across many agents. The company still makes exceptions when there is a specific reason, and Scanlan admits that Intercom uses Codex for code reviews. The broader aim was to make Claude Code capable of acting like a senior engineer on any technical task across Intercom.
The agent had to be onboarded to Intercom's monolith and internal ways of working
Claude Code needed the same context that a new Intercom engineer would receive. That included Rails conventions, architecture, React patterns, testing standards, security rules, and knowledge accumulated during 15 years of software development. Intercom encoded this context in engineering guidance, skills, and hooks. When the agent took the wrong path or hit an issue, the team updated the guidance, creating a feedback loop. The goal covered every part of technical work, including planning, debugging, testing, and production work, with permissions, controls, and audits in place.
Small, tested skills became the durable unit of automation
Scanlan says Intercom focuses on small, high-quality, durable, testable skills rather than asking every engineer to build a custom multi-agent orchestrator. The company uses its history of code changes, incidents, and other work for backtesting and evaluation. It also works to make skills self-updating and continuously improve them. The implementation details may change as tools evolve, but written guidance for how work is done at Intercom should remain useful. This approach lets the company benefit from new software built elsewhere instead of maintaining every capability itself.
Agents work better when engineers give them problems instead of step-by-step tasks
Scanlan's central working principle is to describe the problem and let the agent decide which skills to invoke and what actions to take. He illustrates this with a security incident involving Snowflake table metadata that had been accidentally published to a public GitHub repository. He opened Claude Code, told it to join the relevant Slack channel, and asked it to investigate. Claude Code found and used an internal skill containing Intercom's data-breach policies, downloaded the files, analyzed them, concluded the material was innocuous, and returned the next steps in about two minutes. Scanlan had not known that skill existed.
Intercom used automation and review controls to increase throughput without dropping compliance
After deciding to standardize on Claude Code in December and rolling it out in January, Intercom reached double its previous pull request throughput in less than a year. Scanlan reports that Claude Code produced pull requests in the 90-percent range and that code review had become the current bottleneck. Automatic approval accounted for 17.6% of pull requests. Intercom used backtesting and human labeling to limit this process to safe, simple changes, then worked with auditors on controls for SOC 2, ISO 27001, and HIPAA. Scanlan says the process requires precise auditing and controls, but does not require a human in the loop for those certifications.
The rollout created new operational pressure while making old maintenance work faster
Intercom connected its agent activity to Honeycomb, recorded session transcripts in S3 for analysis, and used the resulting data to understand skill usage and effectiveness. Its CI infrastructure melted under the increase in activity and had to be fixed. Scanlan also describes defects being closed faster, with some teams working toward clearing their backlogs. A skill for fixing flaky Rails specs emerged through repeated feedback rather than being designed in one sitting. It now handles work that Scanlan says would impress him if done by a very senior Rails engineer. Claude Code also spread beyond software engineering, with other teams asking to use the company's internal setup.
"Claude just automatically downloaded the files, did full analysis, concluded it was innocuous, told me all next steps."14:31
Who should watch
Engineering leaders in medium or large organizations who need a concrete plan for rolling out coding agents across hundreds of people.
Platform and developer productivity teams working with a large Rails monolith, internal tooling, recurring maintenance, and production access controls.
Engineers who use agents as command runners and want to learn how reusable skills can handle broader problems.
