How Intercom 2x’d their engineering velocity in 9 months with Claude Code | Brian Scanlan
How I AI
4 DAYS AGO
How Intercom 2x’d their engineering velocity in 9 months with Claude Code | Brian Scanlan
How Intercom 2x’d their engineering velocity in 9 months with Claude Code | Brian Scanlan
How I AI
4 DAYS AGO
This episode features Brian Scanlan, Senior Principal Engineer at Intercom, discussing how the company achieved a dramatic acceleration in engineering velocity and quality by embedding AI deeply into its development culture and infrastructure.
Intercom doubled its merged PRs per R&D employee in nine months by adopting Claude Code organization-wide—not just among engineers, but also designers, PMs, and TPMs. Central to this success was building telemetry infrastructure using Honeycomb, Snowflake, and S3 to measure AI usage, skill effectiveness, and quality outcomes—revealing real-time insights that informed iteration and accountability. The team developed a skills repository with enforceable hooks to standardize engineering practices, exemplified by a self-improving 'flaky spec' skill that scaled from 1x to 100x capability through continuous learning and evaluation. They reimagined workflows as 'agent-first', treating AI spend as strategic investment rather than cost center, and prioritized CLI-first, MCP-ready, and ephemeral API design to make their product—and internal tools—agent-friendly. Crucially, leadership enabled rapid adoption through psychological safety, explicit permission to experiment, and accountability frameworks—transforming backlog zero from aspirational to achievable while freeing engineers to focus on novel, high-impact problems.
00:00
00:00
AI doubles engineering throughput at Intercom in nine months
02:41
02:41
Intercom has met the moment in product and team transformation with AI
05:01
05:01
Engineers using Claude Code during PTO return more skilled, suggesting companies should increase PTO and parental leave
07:03
07:03
The engineering team has twice the throughput compared to nine months ago
12:51
12:51
All technical work will become agent-first
14:28
14:28
Intercom is currently focused on using Opus for everything and treating it as an investment, delaying cost optimization until they've reaped significant benefits
19:27
19:27
AI magnifies both strengths and weaknesses
21:22
21:22
Intercom built an LLM judge to evaluate pull request description quality after Claude Code produced poor outputs
24:03
24:03
Enforcing high-quality PR descriptions via Git hooks increased engineering competence
29:37
29:37
It's great when robots can do work as well as the best engineers
30:16
30:16
Basic usage info is sent to Honeycomb via a shared key to support internal skill discovery
32:10
32:10
Intercom built a tool to provide personalized feedback on users' Claude Code usage compared to others in the organization
36:09
36:09
Upload Claude Code session JSON files to S3 for anonymization and internal evaluation
39:20
39:20
Core plugins have safety hooks and require rigorous evaluation and testing
44:50
44:50
The skill achieves 100x capability through feedback loops and cross-repo reliability
46:45
46:45
Re-implementing a Go microservice in Ruby was accomplished in a single Claude Code session
52:31
52:31
Lobster emojis and live broadcast serve as a growth hack
53:32
53:32
SaaS products must be agent-friendly, supporting CLIs, MCPs, and ephemeral APIs for autonomous agent decision-making
1:01:55
1:01:55
Products don't need to be perfect for agents—they need navigable CLIs and APIs
1:03:49
1:03:49
Escape button usage reveals critical conversion drop-off points
1:11:00
1:11:00
All work will be agent-first in the near future
