Claude Code+75 Strong positive

Operator Insights

What are engineering leaders learning from deploying Claude Code?

24 accounts: 1 failure, incident or rollback, 1 left it or turned it down, 5 switched, 8 run it beside another tool, 1 evaluated it, 5 adopted it and 3 reported by others

Failures, incidents and rollbacks1

Stella LaurenzoEngineer running Claude Code fleetsIREE
Open-source compiler work on IREE. February to April 2026, 6,852 session files mined.
GitHub (anthropics/claude-code issues), published Apr 2, 2026Read the write-up (opens GitHub (anthropics/claude-code issues) in a new tab)
Observation
An engineer running 5 to 10 Claude Code sessions at once rolled back to one after a quality regression.
Takeaway
Parallel sessions multiply a regression as fast as they multiply output. Track merged results per session so a model change shows up before the fleet grows.
What happened
In a GitHub issue of Apr 2, 2026, Laurenzo reported from 6,852 Claude Code session files that February work with one to three sessions produced 191,000 lines of merged code from 1,498 API requests. A quality regression from March forced her to shut the fleet of 5 to 10 or more concurrent sessions down and roll back to a single session.
“1,498 API requests produced 191,000 lines of merged code.”
Stella Laurenzo
- Public practitioner write-up
- Self-reported, measured from session logs
Deployment details
Open-source compiler work on IREE. February to April 2026, 6,852 session files mined.
- Public practitioner write-up
- Self-reported, measured from session logs
Team
One engineer's session fleet
Duration
About 3 months
Workload
IREE compiler projects
Product version
Claude Code
Measurement
Merged lines and API requests mined from 6,852 session files

Left it or turned it down1

Anurag SinghTech journalistXDA Developers
One writer's own coding projects. Six months of daily use, to July 2026.
xda-developers.com, published Jul 27, 2026Read the write-up (opens xda-developers.com in a new tab)
Observation
A daily user cancelled Claude because Claude Code's US$20 limit rarely let a project finish in one session.
Takeaway
On the entry plan, the usage window decides more than model quality. Price heavy users on Max or metered credits before judging the tool.
What happened
After six months of daily Claude Code use on the US$20 Pro plan, Singh cancelled Claude in July 2026. He nearly always ran out of usage before a project was done and moved it to Codex or Antigravity; on the US$20 ChatGPT Plus plan, Codex with GPT-5.6 Sol let him work much longer before the limit.
“I’ve almost never started and completed a project in a single Claude Code session because I nearly always run out of usage and have to wait for the limit to reset. Instead of waiting, I usually end up moving the project to either Codex or Antigravity.”
Anurag Singh
- Public practitioner write-up
- Self-reported
Deployment details
One writer's own coding projects. Six months of daily use, to July 2026.
- Public practitioner write-up
- Self-reported
Team
One writer
Duration
Six months
Workload
Personal coding projects
Product version
Claude Code on the Pro plan, then Codex on ChatGPT Plus
Model
Opus 4.8, then GPT-5.6 Sol

Switched to or from it5

Michael ReddingPrincipal Product ManagerIndeed
Engineering organization of about 2,000 engineers. 2025 to mid-2026.
DX Engineering Enablement podcast, published Jun 29, 2026Listen to the episode (opens DX Engineering Enablement podcast in a new tab)
Observation
Indeed retired Copilot and Cody once its engineers had moved to agents, mostly Claude Code.
Takeaway
Adoption followed the tools engineers chose for themselves. Budget and training did not move usage until the company switched to the agents they already wanted.
What happened
With GitHub Copilot and Sourcegraph Cody, unlimited budgets and training, weekly AI use at Indeed stalled near 25% of developers. Developers began paying for Cursor themselves, Indeed switched to agentic tools, and it retired Copilot and Cody once their use had fallen to about 3%. By the June 2026 episode its mix was mostly Claude Code, with a little Cursor and some Windsurf and Amp, and engineers often ran several at once.
“And if you’re interested, our breakdown rate, we’re mostly Claude Code, a little bit Cursor, some Windsurf, and then down at the bottom, Amp. If you actually add up these percentages, it’s definitely more than 100%. So we also see that our devs are using multiple tools at the same time.”
Michael Redding
- Podcast episode (DX Engineering Enablement, published on YouTube Jun 29, 2026)
- Organization metrics reported
Deployment details
Engineering organization of about 2,000 engineers. 2025 to mid-2026.
- Podcast episode (DX Engineering Enablement, published on YouTube Jun 29, 2026)
- Organization metrics reported
Team
Indeed engineering, about 2,000 engineers
Duration
About 18 months, 2025 to mid-2026
Workload
Product development across Indeed's engineering organization
Product version
Claude Code, with Cursor, Windsurf and Amp
Measurement
Weekly active use of AI tools, about 25% before the switch

Applied IntuitionApr 27, 2026Latent SpaceAt Applied Intuition, Claude Code replaced Cursor as the most used tool.
Peter LudwigCo-founder and chief technology officerApplied Intuition
Company-wide engineering, with an internal adoption leaderboard. To April 2026.
Latent Space, published Apr 27, 2026Listen to the episode (opens Latent Space in a new tab)
Observation
At Applied Intuition, Claude Code replaced Cursor as the most used tool.
Takeaway
Letting engineers choose and measuring what they use shows which tool earns the seats.
What happened
Applied Intuition uses every coding tool and runs an internal leaderboard to encourage adoption. Cursor was the most used tool in the company for a good while, and by April 2026 Claude Code had taken that place.
“Yeah. We use everything. So Cursor was, I think the hottest tool in the company for a good while. Now Claude Code, I think has taken the reign on that. We have a internal leader, leaderboard that we use just to sort of encourage adoption…”
Peter Ludwig, Applied Intuition, at 21:55 in the recording
Podcast episode
Organization usage reported
Deployment details
Company-wide engineering, with an internal adoption leaderboard. To April 2026.
Podcast episode
Organization usage reported
Team
Applied Intuition engineering
Workload
Autonomy and vehicle software
Product version
Claude Code, after Cursor
Measurement
Internal adoption leaderboard
SuperhumanAug 18, 2026Merantix AI CampusSuperhuman moved from Copilot to Cursor to Claude Code, then built its own reviewer on it.
Yaroslav (Yarik)Tech lead for the coding-agent rolloutSuperhuman
Engineering organization at Superhuman, formerly Grammarly. About a year of rollout, to August 2026.
Merantix AI Campus, published Aug 18, 2026Watch the talk (opens Merantix AI Campus in a new tab)
Observation
Superhuman moved from Copilot to Cursor to Claude Code, then built its own reviewer on it.
Takeaway
A team that builds its own review layer on the agent keeps control of what review checks. Plan for the upkeep that choice brings.
What happened
Superhuman's engineers started with Copilot's autocomplete, moved to Cursor, and switched to Claude Code after Anthropic offered an early beta about 11 months before the talk. The tech lead says he has not written code in the editor since. Its internal code reviewer runs a Claude Code coordinator over a swarm of review agents and calls Codex for a second opinion on complex pull requests, kept in place of Anthropic's Code Review because Superhuman can tailor it.
“11 months ago to Superhuman reached out Anthropic reached to us and offered to do the beta of this new cloud cloud tool that basically changed everything overnight, at least for me. I was like the first participant in this POC. Since then, haven't written a single line of code in in the editor.”
Yaroslav (Yarik), Superhuman, at 3:58, from automatic captions
Conference talk
Self-reported
Deployment details
Engineering organization at Superhuman, formerly Grammarly. About a year of rollout, to August 2026.
Conference talk
Self-reported
Team
Superhuman engineering
Duration
About 11 months on Claude Code
Workload
Product engineering and internal code review
Product version
Claude Code, with Codex as second reviewer
EverySep 22, 2026every.toTwo Every staff who switched from Claude Code to Codex were reconsidering after Opus 5.5.
Katie ParrottStaff writerEvery
Every's staff across evals, design and leadership. Account published September 22, 2026.
every.to, published Sep 22, 2026Read the write-up (opens every.to in a new tab)
Observation
Two Every staff who switched from Claude Code to Codex were reconsidering after Opus 5.5.
Takeaway
Preference between the two follows the newest model, so a team that buys one should expect to revisit it at each release.
What happened
Every's head of evals Mike Taylor and designer Tyler Nishida switched from Claude Code to Codex in the months before September 22, 2026, when Every wrote that Opus 5.5 had them reconsidering that switch. Taylor still prefers the Codex harness but would take a worse harness for a better model, and CEO Dan Shipper still splits his use 80/20 toward Codex while spending more tokens on Claude.
“I prefer the Codex harness, but I’d rather have a worse harness if I can work with a better model.”
Mike Taylor, Every
Public organization write-up
Self-reported
Deployment details
Every's staff across evals, design and leadership. Account published September 22, 2026.
Public organization write-up
Self-reported
Team
Every staff
Duration
Several months on Codex before September 22, 2026
Workload
Product and editorial engineering
Product version
Claude Code and Codex
Model
Opus 5.5
SanitySep 2, 2025Sanity EngineeringAt Sanity, most tasks took three attempts, and the engineer stayed responsible for what shipped.
Vincent QuigleyThen Staff EngineerSanity
Individual engineer on a product team. 6 weeks. Production SaaS.
Sanity Engineering, published Sep 2, 2025Read the write-up (opens Sanity Engineering in a new tab)
Observation
At Sanity, most tasks took three attempts, and the engineer stayed responsible for what shipped.
Takeaway
Expect to discard most first attempts. Review agent output in stages, and keep the engineer who ran the session accountable for what ships.
What happened
After 18 months with Cursor, Quigley moved to Claude Code, running several sessions in parallel with CLAUDE.md and Linear, Notion, and GitHub as context. In the write-up, the first attempt at a feature is about 95% unusable and the third is workable. Quigley estimates shipping features two to three times faster and puts the budget for a senior engineer using it heavily at US$1,000 to US$1,500 a month.
“The ROI is obvious, but budget for $1000-1500/month for a senior engineer going all-in on AI development. It's also reasonable to expect engineers to get more efficient with AI spend as they get good with it, but give them time.”
Vincent Quigley
Public engineering write-up
Self-reported estimates
Deployment details
Individual engineer on a product team. 6 weeks. Production SaaS.
Public engineering write-up
Self-reported estimates
Team
One staff engineer at Sanity
Duration
6 weeks
Workload
Production SaaS features

Run it beside another tool8

StripeFeb 19, 2026Stripe Dot Dev BlogStripe syncs one set of rules into Cursor, Claude Code and its own agent.
Alistair GrayEngineerStripe
Stripe engineering, February 2026.
Stripe Dot Dev Blog, published Feb 19, 2026Read the write-up (opens Stripe Dot Dev Blog in a new tab)
Observation
Stripe syncs one set of rules into Cursor, Claude Code and its own agent.
Takeaway
Write project guidance once and sync it to every agent, so engineers can move between tools without losing conventions.
What happened
Stripe makes Cursor and Claude Code available for supervised work and runs its own unattended agent, minions, beside them for one-shot tasks; the three are its most used coding agents. Stripe syncs its Cursor rules into a format Claude Code reads so all three share the same guidance, and over 1,300 merged pull requests a week are written entirely by minions.
“We also now sync our Cursor rules into a format that Claude Code can read as well, so that our three most popular coding agents (minions, Cursor, and Claude Code) can all benefit from the guidance that lives in rule files that Stripe engineers are scaffolding in our codebase.”
Alistair Gray
Public engineering write-up
Organization metrics reported
Deployment details
Stripe engineering, February 2026.
Public engineering write-up
Organization metrics reported
Team
Stripe engineering
Workload
Stripe's codebase
Product version
Claude Code, Cursor and Stripe's minions agent
Measurement
Merged pull requests a week written by minions
ShopifyApr 22, 2026Latent SpaceAt Shopify, command-line agents such as Claude Code grew fastest once employees could choose.
Mikhail ParakhinChief technology officerShopify
Every Shopify employee may choose a coding tool. December 2025 to April 2026.
Latent Space, published Apr 22, 2026Listen to the episode (opens Latent Space in a new tab)
Observation
At Shopify, command-line agents such as Claude Code grew fastest once employees could choose.
Takeaway
Usage data from open choice shows where the work is moving; Shopify reads it before buying more seats of any one tool.
What happened
Shopify's internal usage chart shows command-line agents such as Claude Code, Codex, Pi and Shopify's internal agent River taking off since December 2025, while IDE tools such as GitHub Copilot and Cursor are not shrinking but grow more slowly.
“The other thing I would claim you could see is that, uh, CLI-based tools and tools that don’t require you to look at the code becoming more popular, and you could see, yeah, various versions of, uh, Cloud Code and Codex and Pi and internal development tools taking off.”
Mikhail Parakhin, Shopify, at 3:52 in the recording
Podcast episode
Organization usage reported
Deployment details
Every Shopify employee may choose a coding tool. December 2025 to April 2026.
Podcast episode
Organization usage reported
Team
Shopify, company-wide
Duration
About 5 months
Workload
Shopify's codebase
Product version
Claude Code beside Codex, Pi, River, Copilot and Cursor
Measurement
Internal usage by tool
BrexJan 17, 2026Latent SpaceBrex buys a few seats of several tools and lets usage decide renewals.
James ReggioChief technology officerBrex
Brex engineering, January 2026.
Latent Space, published Jan 17, 2026Listen to the episode (opens Latent Space in a new tab)
Observation
Brex buys a few seats of several tools and lets usage decide renewals.
Takeaway
Small seat counts across tools, with usage read at renewal, avoid locking a company into one vendor while the tools change.
What happened
Brex procures a small number of seats of several coding tools and lets employees pick Cursor, Windsurf or Claude Code credits, using usage trends to resist company-wide contracts at renewal. Some engineers stay on Claude Code after nine months because they know it.
“…some folks are just like, I want to stick with Cloud Code because I know it now. I’ve been working with it for like nine months, so I don’t need to keep switching.”
James Reggio, Brex, at 21:15 in the recording
Podcast episode
Self-reported
Deployment details
Brex engineering, January 2026.
Podcast episode
Self-reported
Team
Brex engineering
Duration
Nine months on Claude Code for some engineers
Workload
Brex's codebase
Product version
Claude Code credits beside Cursor and Windsurf
Measurement
Usage trends by tool at renewal
SalesforceDec 16, 2025Salesforce EngineeringSalesforce runs Claude Code beside Cursor and Windsurf, each for its own workflow.
Darryn DiekenChief Availability OfficerSalesforce
Salesforce engineering, December 2025.
Salesforce Engineering, published Dec 16, 2025Read the write-up (opens Salesforce Engineering in a new tab)
Observation
Salesforce runs Claude Code beside Cursor and Windsurf, each for its own workflow.
Takeaway
Assign each tool the workflow it suits, then scale review and deployment checks before output rises.
What happened
Salesforce runs a portfolio of coding tools: Cursor and Windsurf for code and test generation, and Claude Code as a command-line tool for automation and scripting, with engineers choosing by workflow. It reports over 85% of eligible engineers using the tools weekly and code in production up 30%, which put pressure on review and deployment.
“AI accelerated development speed, increasing code in production by 30%. This surge created significant pressure on downstream processes.”
Darryn Dieken
Public engineering write-up
Organization metrics reported
Deployment details
Salesforce engineering, December 2025.
Public engineering write-up
Organization metrics reported
Team
Salesforce engineering
Workload
Salesforce products
Product version
Claude Code, Cursor and Windsurf
Measurement
Weekly use among eligible engineers, code in production
37signalsAug 26, 2026Lex Fridman PodcastAt 37signals, Claude Code does the work and Codex reviews every job.
DHHChief technology officer37signals
DHH's own development on Omarchy and 37signals work, August 2026.
Lex Fridman Podcast, published Aug 26, 2026Listen to the episode (opens Lex Fridman Podcast in a new tab)
Observation
At 37signals, Claude Code does the work and Codex reviews every job.
Takeaway
Two agents from different model families catch each other's mistakes. Budget a reviewing agent beside the one that writes the code.
What happened
DHH drives his agent work in Claude Code, which he calls the best harness for running several agents, and ends every job with a Codex review at xhigh effort, his standard procedure. He often runs Claude in one pane and Codex below it, and after his agents push to GitHub, Copilot's pull request review keeps finding things that are broken.
“I’ve mainly been driving with Claude, which is actually interesting because I have some other reservations about Anthropic. But the reason I’m sticking with Claude is, in my opinion, they actually have the best harness.”
DHH, at 2:39:24 in the recording
Podcast episode
Self-reported
Deployment details
DHH's own development on Omarchy and 37signals work, August 2026.
Podcast episode
Self-reported
Team
DHH's own development
Workload
Omarchy and 37signals code
Product version
Claude Code, with Codex and GitHub Copilot review
Model
Opus 5 or Fable writing, Codex at xhigh effort reviewing
DatabricksJun 24, 2026Latent SpaceDatabricks built one internal wrapper over Claude Code and Codex.
Matei ZahariaCo-founderDatabricks
Databricks developer infrastructure, June 2026.
Latent Space, published Jun 24, 2026Listen to the episode (opens Latent Space in a new tab)
Observation
Databricks built one internal wrapper over Claude Code and Codex.
Takeaway
A thin internal layer over both agents keeps the choice of model open without two rollouts.
What happened
Databricks' developer infrastructure team built Isaac, a wrapper on Claude Code and Codex that lets engineers run either agent in web sandboxes or on their own machines.
“…they built something called Isaac, that’s like a wrapper on Claude Code and Codex, and, lets you use them either on the web in, like, sandboxes or, just on your dev machine or on your laptop or whatever.”
Matei Zaharia, Databricks, at 2:22 in the recording
Podcast episode
Self-reported
Deployment details
Databricks developer infrastructure, June 2026.
Podcast episode
Self-reported
Team
Databricks engineering
Workload
Databricks product development
Product version
Claude Code and Codex behind Isaac
CoinbaseJan 6, 2026AI Native Dev (Tessl)A Coinbase engineer works in Cursor and hands deep tasks to Claude Code.
Sepehr KhosraviMachine learning platform engineerCoinbase
One platform engineer at Coinbase, which tracks AI use per developer. January 2026.
AI Native Dev (Tessl), published Jan 6, 2026Listen to the episode (opens AI Native Dev (Tessl) in a new tab)
Observation
A Coinbase engineer works in Cursor and hands deep tasks to Claude Code.
Takeaway
Keep the editor for fast, visible edits and the terminal agent for the hard task, and expect the second to cost more tokens.
What happened
Khosravi runs Cursor with Claude Code in the terminal beside it and does 80 to 90% of his work in Cursor. He starts deep, difficult tasks in Claude Code, which takes more tokens and usually gives a better engineered answer; on one work project it searched the web and open-source repositories before proposing a solution.
“Cursor's good cuz it gives you the visual ID. You can see what you're doing. It's faster than Claude Code is, uses less tokens, and you can switch between different AI models. But if you're trying to go deep on a task that's difficult, Claude Code does a lot more thinking, takes up a lot more tokens, but usually gives you a better engineered answer.”
Sepehr Khosravi, Coinbase, at 12:09, from automatic captions
Podcast episode
Self-reported
Deployment details
One platform engineer at Coinbase, which tracks AI use per developer. January 2026.
Podcast episode
Self-reported
Team
One engineer
Workload
Machine learning platform work
Product version
Claude Code beside Cursor
BetterUpNov 24, 2025AI EngineerBetterUp's AI enablement lead keeps Claude Code as his daily driver and added Codex to avoid lock-in.
Brian JohnPrincipal full-stack engineer, AI enablement for R&DBetterUp
BetterUp R&D AI enablement, November 2025.
AI Engineer, published Nov 24, 2025Watch the talk (opens AI Engineer in a new tab)
Observation
BetterUp's AI enablement lead keeps Claude Code as his daily driver and added Codex to avoid lock-in.
Takeaway
Keep workflows portable across agents, and expect the second agent to need its own setup.
What happened
John has used Claude Code as his daily driver since its release and added the Codex CLI beside it so he is not locked in to one tool or one model family. He built a sub-agent workaround for Codex, found its sandbox resisted it, and finds Codex slower and built for hands-off work.
“I've been using Clog Code as my daily driver since the very beginning. It's a great tool. It's got tons of bells and whistles. It's got great models, and I use sub agents all the time. But I don't want to be locked in to one tool, and I really don't want to be locked in to one model family.”
Brian John, BetterUp, at 0:59, from automatic captions
Conference talk
Self-reported
Deployment details
BetterUp R&D AI enablement, November 2025.
Conference talk
Self-reported
Team
BetterUp R&D
Workload
AI enablement across R&D
Product version
Claude Code, with the Codex CLI

Evaluated it1

Prime IntellectSep 26, 2026AI EngineerIn Prime Intellect's research race, Claude Code beat the human record by 50 to 60 steps, and sat idle a third of the time.
Elie BakouchResearch engineerPrime Intellect
Automated research on Prime Intellect's cluster, about July 2026.
AI Engineer, published Sep 26, 2026Watch the talk (opens AI Engineer in a new tab)
Observation
In Prime Intellect's research race, Claude Code beat the human record by 50 to 60 steps, and sat idle a third of the time.
Takeaway
For unattended runs, watch idle time as closely as results, because an agent that stops to wait needs monitoring.
What happened
Prime Intellect evaluated Claude Code on Opus 4.8 against Codex on GPT-5.5 as autonomous agents on a public optimizer speedrun. Bakouch reports that Claude Code beat the human record of about 2,990 steps by 50 to 60 steps; Claude Code sat idle about a third of the time, while Codex worked almost continuously and burned more tokens.
“I think at the time the best uh record was like uh 2,990 step and we beat it by like uh uh 50 or 60 step for code and codeex was like 20 step above.”
Elie Bakouch, Prime Intellect, at 11:13, from automatic captions
Conference talk
Measured on one task
Deployment details
Automated research on Prime Intellect's cluster, about July 2026.
Conference talk
Measured on one task
Team
Prime Intellect research
Duration
One speedrun
Workload
Optimizer speedrun
Product version
Claude Code and Codex
Model
Opus 4.8 and GPT-5.5
Measurement
Steps to the target loss, idle time

Adopted it5

SpotifyJun 3, 2026Spotify EngineeringAt Spotify, the constraint moved from writing code to reviewing it.
Niklas GustavssonChief Architect and VP of EngineeringSpotify
Engineering organization working across the backend fleet and a client monorepo.
Spotify Engineering, published Jun 3, 2026Read the write-up (opens Spotify Engineering in a new tab)
Observation
At Spotify, the constraint moved from writing code to reviewing it.
Takeaway
Consistent codebases and platform context make agents more effective, and every gain in output adds pull requests someone has to review.
What happened
Spotify runs Claude through the Agent SDK inside Honk, its background coding agent, for fleet-wide migrations; its most recent Java migration across backend services took three days. More than 99% of Spotify engineers use AI coding tools every week, and pull request frequency is up 76%, all of it work to review.
“The flip side: we now have 76% more PRs to review.”
Spotify Engineering
Public engineering write-up of a conference talk
Organization metrics reported
Deployment details
Engineering organization working across the backend fleet and a client monorepo.
Public engineering write-up of a conference talk
Organization metrics reported
Team
Spotify engineering
Workload
Backend fleet migrations and client monorepo
Product version
Claude Code and the Claude Agent SDK
Model
Opus 4.5, the release Spotify ties to its adoption jump
Measurement
Weekly tool usage, productivity survey, pull request frequency
MercariJun 22, 2026Mercari EngineeringA Mercari iOS engineer ran 22 Claude Code sessions to migrate a design system.
cyaniOS engineer, Mercoin client teamMercari
Mercoin iOS client team, which uses Claude Code for migrations. June 2026.
Mercari Engineering, published Jun 22, 2026Read the write-up (opens Mercari Engineering in a new tab)
Observation
A Mercari iOS engineer ran 22 Claude Code sessions to migrate a design system.
Takeaway
A repetitive migration splits cleanly across parallel sessions when the commands and context are shared.
What happened
An iOS engineer on Mercari's Mercoin client team ran 22 Claude Code sessions across git worktrees, with shared slash commands kept in an Obsidian vault, to migrate app screens from DesignSystem V3 to V4. The migration produced around 140 pull requests.
“Over the course of this migration, I have created around 140 Pull Requests (PRs)…”
cyan, Mercari
Public engineering write-up
Self-reported
Deployment details
Mercoin iOS client team, which uses Claude Code for migrations. June 2026.
Public engineering write-up
Self-reported
Team
One iOS engineer
Workload
DesignSystem V3 to V4 screen migration
Product version
Claude Code across git worktrees
Measurement
Pull requests created, around 140
incident.ioJun 27, 2025incident.io blogincident.io went from no Claude Code to four or five agents in parallel within four months.
Rory BainProduct Engineerincident.io
incident.io product engineering, to June 2025.
incident.io blog, published Jun 27, 2025Read the write-up (opens incident.io blog in a new tab)
Observation
incident.io went from no Claude Code to four or five agents in parallel within four months.
Takeaway
Git worktrees let several agents work on separate features without colliding.
What happened
Within four months of Claude Code's launch, incident.io went from not using it to running four or five Claude Code agents at once, each on a separate git worktree. One session costing about US$8 in Claude credit produced a fix that cut its API generation build step by 18%, or 30 seconds.
“We've gone from no Claude Code to simultaneously running four or five Claude agents, each working on different features in parallel.”
Rory Bain
Public engineering write-up
Self-reported
Deployment details
incident.io product engineering, to June 2025.
Public engineering write-up
Self-reported
Team
incident.io product engineering
Duration
Four months
Workload
Product features and build tooling
Product version
Claude Code on git worktrees
Measurement
Build step time, cut 18%
IntercomJun 15, 2026DX on YouTubeIntercom, now Fin, tripled engineering throughput after treating Claude Code as a platform.
Brian ScanlanSenior Principal Systems EngineerIntercom
R&D organization. 16 months. Monolithic Rails applications.
DX on YouTube, published Jun 15, 2026Watch the talk (opens DX on YouTube in a new tab)
Observation
Intercom, now Fin, tripled engineering throughput after treating Claude Code as a platform.
Takeaway
Scanlan credits the gain to skills and context the whole organization maintains, and to a stated expectation that every engineer uses agents.
What happened
Intercom set a goal of doubling merged pull requests per R&D head, standardized on Claude Code, and built shared skills, plugins, and telemetry around it. Throughput doubled in nine months and tripled in sixteen. At the time of the talk, 95.9% of pull requests were authored by Claude, and about 17% were approved by Claude with no human reviewer.
Conference talk
Organization metrics reported
Deployment details
R&D organization. 16 months. Monolithic Rails applications.
Conference talk
Organization metrics reported
Team
Intercom R&D
Duration
16 months
Workload
Monolithic Rails applications with React front ends
Product version
Claude Code with internal plugins and skills
Measurement
Merged pull requests per R&D head, share of agent-approved pull requests, defect backlog
KnakMay 7, 2026Knak BlogKnak's designers now prototype in the production codebase with Claude Code.
Jack SteeleDirector of EngineeringKnak
Knak product design and engineering, May 2026.
Knak Blog, published May 7, 2026Read the write-up (opens Knak Blog in a new tab)
Observation
Knak's designers now prototype in the production codebase with Claude Code.
Takeaway
Opening the codebase to designers takes engineering work first: guardrails and setup before access.
What happened
A tiger team of Knak developers spent six weeks making the production repository safe for designers, who now prototype directly in it using Claude Code in place of Figma and Lovable handoffs. Knak gives no numeric outcome.
“By letting designers prototype directly in their production codebase with Claude Code, Knak replaced fragmented design handoffs with a single shared workflow.”
Jack Steele
Public engineering write-up
No outcome measured
Deployment details
Knak product design and engineering, May 2026.
Public engineering write-up
No outcome measured
Team
Knak designers and a developer tiger team
Duration
Six weeks of setup
Workload
Prototyping in the production codebase
Product version
Claude Code

Reported by others3

Natalie LungJun 3, 2026Simon Willison's WeblogUber capped agent spend at US$1,500 per employee a month after using up its 2026 AI budget.
Natalie LungReporter, Bloomberg News, quoted by Simon WillisonBloomberg
Every Uber employee using agentic coding tools, 2026.
Simon Willison's Weblog, published Jun 3, 2026Read the report (opens Simon Willison's Weblog in a new tab)
Observation
Uber capped agent spend at US$1,500 per employee a month after using up its 2026 AI budget.
Takeaway
Set per-member spend limits before usage grows, because agent spend compounds faster than seat budgets plan for.
What happened
Simon Willison's weblog quoted Bloomberg News on Jun 3, 2026: after Uber spent its 2026 AI budget in four months, it limited every employee to US$1,500 of monthly token spending per agentic coding tool, a cap that names Cursor and Claude Code.
“The rideshare giant is limiting all employees to $1,500 in monthly token spending per AI coding tool…”
Bloomberg News, quoted by Simon Willison
Second-hand news report (Simon Willison's weblog, quoting Bloomberg News)
Spokesperson statement reported
Deployment details
Every Uber employee using agentic coding tools, 2026.
Second-hand news report (Simon Willison's weblog, quoting Bloomberg News)
Spokesperson statement reported
Team
Uber, all employees
Workload
Agentic coding across Uber
Product version
Claude Code and Cursor
Measurement
Monthly token spend per tool
Uday Kiran MedisettyOct 14, 2026The Pragmatic EngineerUber supports Claude Code, Codex and OpenCode behind one wrapper that carries each engineer's setup.
Uday Kiran MedisettyDistinguished EngineerUber
Uber's AI Foundations developer platform, October 2026.
The Pragmatic Engineer, published Oct 14, 2026Read the report (opens The Pragmatic Engineer in a new tab)
Observation
Uber supports Claude Code, Codex and OpenCode behind one wrapper that carries each engineer's setup.
Takeaway
Keep skills and settings in a layer the company owns, so switching agents costs engineers nothing.
What happened
The Pragmatic Engineer reported on Oct 14, 2026 that Uber supports Claude Code, Codex and OpenCode as harnesses. AIFX, an internal wrapper over all three, moves an engineer's settings, skills and plugins when they switch harness.
“Uber currently supports Claude Code, Codex, and OpenCode as harnesses. AIFX (short for AI Foundations) is a wrapper on top of all of these, so when an engineer changes harnesses, all their settings, skills and plugins move with them.”
The Pragmatic Engineer
Second-hand news report (The Pragmatic Engineer)
Organization platform described
Deployment details
Uber's AI Foundations developer platform, October 2026.
Second-hand news report (The Pragmatic Engineer)
Organization platform described
Team
Uber engineering
Workload
Uber's developer platform
Product version
Claude Code, Codex and OpenCode behind AIFX
James FarisJul 29, 2026businessinsider.comDisney is dropping Copilot and keeping Claude Enterprise, Business Insider reported.
James FarisReporterBusiness Insider
Disney's US technology organization, August 2026.
businessinsider.com, published Jul 29, 2026Read the report (opens businessinsider.com in a new tab)
Observation
Disney is dropping Copilot and keeping Claude Enterprise, Business Insider reported.
Takeaway
Large buyers are consolidating on the agents engineers already use, and retiring the assistants they ignore.
What happened
Business Insider reported on Jul 29, 2026, from an internal message, that Disney is dropping GitHub Copilot, Kiro and Amazon Q for US staff in August, preparing to adopt Codex, and keeping Claude Enterprise and Cursor. One senior engineer said about 80% of their AI use is Claude at the terminal.
“One high-level software engineer said their AI tool usage was "probably around 80% Claude at the terminal when coding."…”
A Disney software engineer, quoted by Business Insider
Second-hand news report (Business Insider, from a Disney internal message)
Employee accounts reported
Deployment details
Disney's US technology organization, August 2026.
Second-hand news report (Business Insider, from a Disney internal message)
Employee accounts reported
Team
Disney technology staff, US
Workload
Software development across Disney
Product version
Claude Enterprise and Cursor kept, Codex added

Failures, incidents and rollbacks1

An engineer running 5 to 10 Claude Code sessions at once rolled back to one after a quality regression.

Left it or turned it down1

A daily user cancelled Claude because Claude Code's US$20 limit rarely let a project finish in one session.

Switched to or from it5

Indeed retired Copilot and Cody once its engineers had moved to agents, mostly Claude Code.

At Applied Intuition, Claude Code replaced Cursor as the most used tool.

Superhuman moved from Copilot to Cursor to Claude Code, then built its own reviewer on it.

Two Every staff who switched from Claude Code to Codex were reconsidering after Opus 5.5.

At Sanity, most tasks took three attempts, and the engineer stayed responsible for what shipped.

Run it beside another tool8

Stripe syncs one set of rules into Cursor, Claude Code and its own agent.

At Shopify, command-line agents such as Claude Code grew fastest once employees could choose.

Brex buys a few seats of several tools and lets usage decide renewals.

Salesforce runs Claude Code beside Cursor and Windsurf, each for its own workflow.

At 37signals, Claude Code does the work and Codex reviews every job.

Databricks built one internal wrapper over Claude Code and Codex.

A Coinbase engineer works in Cursor and hands deep tasks to Claude Code.

BetterUp's AI enablement lead keeps Claude Code as his daily driver and added Codex to avoid lock-in.

Evaluated it1

In Prime Intellect's research race, Claude Code beat the human record by 50 to 60 steps, and sat idle a third of the time.

Adopted it5

At Spotify, the constraint moved from writing code to reviewing it.

A Mercari iOS engineer ran 22 Claude Code sessions to migrate a design system.

incident.io went from no Claude Code to four or five agents in parallel within four months.

Intercom, now Fin, tripled engineering throughput after treating Claude Code as a platform.

Knak's designers now prototype in the production codebase with Claude Code.

Reported by others3

Uber capped agent spend at US$1,500 per employee a month after using up its 2026 AI budget.

Uber supports Claude Code, Codex and OpenCode behind one wrapper that carries each engineer's setup.

Disney is dropping Copilot and keeping Claude Enterprise, Business Insider reported.