Scaling Engineering Velocity through Agentic AI: Intercom's Framework
An analysis of how Intercom doubled its R&D throughput by adopting an agent-first engineering culture. The discussion focuses on the 'Software Factory' model, telemetry-driven AI adoption, and the transition toward agent-friendly SaaS architectures.
Unlocking Exponential R&D Throughput
For most leadership teams, increasing engineering velocity is a constant struggle of balancing feature delivery with technical debt. However, Intercom has demonstrated that it is possible to double R&D throughput—measured by merged Pull Requests (PRs) per head—within nine months by moving beyond simple AI autocomplete toward an "agent-first" organizational model.
The 'Software Factory' Approach
Rather than treating AI as a peripheral tool, the goal is to build a "Software Factory." This involves creating a deterministic pipeline where organizational standards are encoded into "Skills" and "Hooks." For example, by implementing a specific CreatePR skill, a company can ensure that AI-generated PR descriptions focus on intent and context rather than just summarizing code, maintaining a high quality bar even as volume increases.
Telemetry-Driven Adoption
Successful AI integration requires treating the engineering organization as a product. This means implementing rigorous telemetry to track how AI skills are invoked and analyzing session data to identify where developers are struggling. By monitoring the "dropout rate" of AI sessions and the effectiveness of specific skills, leadership can iteratively optimize the developer experience and remove systemic bottlenecks.
The Shift to Agent-Friendly SaaS
As AI agents begin to handle more software installation and configuration, the traditional UI/UX paradigm is shifting. The emerging requirement is for "Agent-friendly" interfaces—robust CLIs, MCPs, and ephemeral APIs—that allow agents to traverse a product's onboarding and setup flow without human intervention.
Conclusion
AI transforms the constraints of software engineering from physical typing and coordination limits to the limits of imagination. By investing in code quality via AI "speed-runs" of tech debt and building a high-trust culture that permits rapid experimentation, organizations can achieve a level of velocity previously thought impossible.
Key insights
-
Engineering throughput can be doubled by transitioning to an "agent-first" mindset where all technical work is reimagined around AI agents rather than just augmenting human typing.
Impact: Dramatic reduction in time-to-market for new features and a significant increase in R&D efficiency.
-
The 'Software Factory' model uses custom 'Skills' and 'Hooks' to enforce deterministic quality standards, preventing the 'slop' often associated with high-volume AI code generation.
Impact: Allows for massive scaling of code output without a corresponding drop in maintainability or architectural integrity.
-
SaaS products must evolve to be 'Agent-friendly,' prioritizing CLIs and APIs over traditional GUIs, as AI agents increasingly become the primary installers and users of software.
Impact: A fundamental shift in customer acquisition and onboarding funnels, where the 'user' is an AI agent.
-
AI makes the resolution of long-standing technical debt and 'flaky' test suites tractable by automating the research and propagation of fixes across large codebases.
Impact: Improved codebase stability and a superior developer experience by eliminating low-value manual maintenance.
Action items
-
Implement event-level telemetry (e.g., via Honeycomb or Snowflake) on all internal AI tools to track skill invocation and identify adoption bottlenecks.
Impact: Enables data-driven optimization of the AI toolchain and provides visibility into real-world developer productivity.
-
Initiate a one-month 'tech debt speed-run' using AI agents to fix known codebase frustrations and flaky tests.
Impact: Rapidly improves codebase health and acts as a catalyst to 'AI-pill' skeptical engineering teams by showing immediate value.
-
Develop an 'Agent-friendly' interface (CLI or MCP) for the product's onboarding and configuration flow to reduce agent-driven installation drop-offs.
Impact: Increases adoption rates as AI agents can seamlessly integrate the product into a customer's ecosystem.
Quotes
“your imagination is now the barrier not the tool”
“We believe that all technical work will become agent first.”
“I often advise a lot of ctos and vps of engineering when figuring out how to get their engineering team ai pilled say everything you hate about the code base go spend a month fixing and see how fast we can speed run that”