Key findings
- Fin, formerly Intercom, reports that merged pull requests per R&D head doubled in nine months after it built a shared library of Claude Code skills and plugins.
- Spotify counts its 76% rise in pull request frequency as that many more pull requests to review.
- Individual engineers describe a learning curve of several attempts per task and a messy first month.
- In METR's randomized trial, experienced developers took 19% longer with AI in early 2025. Its late-2025 follow-up estimated speedups of 4% to 18%, which METR calls an unreliable signal.
Teams that built shared context and tooling around a coding agent report the largest gains, and then review becomes their constraint. Fin (formerly Intercom), Spotify, Sanity, incident.io, Expo, and Puzzmo have published how agents changed their work, and METR's randomized studies measure smaller effects than those teams report.
Shared context drives the gains
Intercom, which renamed itself Fin in May 2026 and has been owned by Salesforce since September 2026, standardized on Claude Code and built a shared library of skills and plugins with telemetry. It reports that merged pull requests per R&D head doubled in nine months and tripled in sixteen (Brian Scanlan, DX talk, Jun 2026).
Spotify runs Claude through the Agent SDK in Honk, its background coding agent. It credits consistent codebases, and the platform context it gives the agent, for better output. Its most recent Java migration across backend services took three days (Spotify Engineering, Jun 2026).
Review is the new bottleneck
Spotify's September 2026 quality review found merged changes more than doubled year over year. The review did not identify AI-authored code as a material direct cause of incidents, and it found change volume outpacing verification controls.
Adoption has a learning curve
A Sanity staff engineer reports that the first attempt at a task is mostly wrong and the third is workable. They estimate two to three times faster shipping and hold the human engineer responsible for every line (Sanity, Sep 2025).
incident.io runs four or five agents in parallel in separate Git worktrees (Jun 2025). Expo's web team found the agent strongest on well-specified tasks in codebases with strong patterns and weaker as context grew (Feb 2026). At Puzzmo, a month and a half of migration and maintenance work did not show clearly in pull request, commit, or line-count charts (Jul 2025).
What the measured evidence says
Most published gains are measured by the team that reports them. METR's randomized trial covered 16 experienced developers on 246 issues in their own repositories. Its 2026 follow-up drew on 57 developers, and METR discounts that newer data because developers increasingly declined to work without AI.
Methodology
GILD drew this from public engineering write-ups, a conference talk, and published research, each linked below with the date GILD checked it. Every figure is its source's own.
Products in this piece
Sources
- Public source
- 10
- Brian Scanlan (Intercom): Doubling the productivity of your engineering team using AI, DX talk (Jun 15, 2026)
- Intercom: How we use Claude Code today at Intercom (Mar 19, 2026)
- Spotify Engineering: Coding Is No Longer the Constraint (Jun 3, 2026)
- Spotify Engineering: AI Changed How Spotify Builds (Sep 16, 2026)
- Sanity: First attempt will be 95% garbage (Sep 2, 2025)
- incident.io: How we're shipping faster with Claude Code and Git Worktrees (Jun 27, 2025)
- Expo: What our web team learned using Claude Code for a month (Feb 6, 2026)
- Puzzmo: 6 Weeks of Claude Code (Jul 30, 2025)
- METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (Jul 10, 2025)
- METR: We are Changing our Developer Productivity Experiment Design (Feb 24, 2026)