--- title: "Your coding-agent adoption rate is not the number your board should be watching" slug: harness-coding-agent-adoption-rate-wrong-board-number date: 2026-06-29 excerpt: "The board keeps asking for your coding-agent adoption rate. A new look inside Shopify and OpenAI suggests that number measures the smallest part of the return, and a better one is hiding in whether the agent's work is visible." featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1782738505481-harness-coding-agent-adoption-rate-wrong-board-number.webp" featured_image_alt: "A factory shop floor reimagined as an open-plan engineering office, where one worker's task is visible to everyone around them, illustrating shared organizational learning from AI agents." canonical_url: https://cerevisor.com/blog/harness-coding-agent-adoption-rate-wrong-board-number updated_at: 2026-06-29T13:08:26.819539+00:00 --- # Your coding-agent adoption rate is not the number your board should be watching A board member asked a founder I know last week what their [coding-agent adoption](/blog/harness-adoption-stops-being-free-june-1) rate was. The founder said 71 percent. The board member nodded, wrote it down, and moved on. Everyone in the room felt good. And I think that number told them almost nothing about whether the company is actually getting anything back. TLDR Adoption rate measures how many engineers have the agent turned on, which is the cheapest part of the return to capture and the easiest to fake. A piece this week on how Shopify and OpenAI run their agents points to a better board number: whether the agent's work is visible enough that one engineer's discovery becomes everyone's. Private adoption at 71 percent can be worth less than visible adoption at 40. ## The seat count became the metric everyone reports to the board Here is the belief, stated plainly, because it is everywhere right now. The percentage of engineers actively using a coding agent is the headline number that proves the investment is working. Higher adoption equals more value. Get the rate up, and the return follows. It is a clean story. It fits on one slide. The vendor dashboard hands it over for free. And after a year of pilots that went nowhere, a high adoption rate feels like proof that this time the thing actually stuck. I understand why it became the default. We measure what is easy to measure, and seat activation is the easiest thing in the building to count. --- ## Why a high adoption rate feels like proof of value The logic is not stupid. For most of software history, getting people to use a new tool was the hard part. Mandate a process, half the team quietly ignores it. So when 71 percent of engineers open the agent every day without being forced, that genuinely is a signal. It means the tool cleared the bar that kills most internal rollouts, which is that people would rather not change how they work. There is also real money behind the number now. Since [metered billing](/blog/weekly-recap-2026-06-05) went live across the major harnesses this month, every active engineer is a variable cost. So the adoption rate doubles as a utilization rate. When the bill runs per seat and per token, idle licenses are pure waste, and a busy one at least looks like it is earning its keep. So adoption rate carries two true meanings at once. It says the team accepted the tool, and it says the spend is being consumed. Both are real. Neither one says the work got better. > A busy seat is not the same as a return. It is just a seat that is busy. ## What Shopify and OpenAI reveal about where the value actually lives Scott Carey had a piece in InfoWorld on June 29 that reframed this better than anything I have read this year. The argument is simple and it lands hard: the value of a coding agent compounds when its work is visible, not when its seat is active. The example is Shopify’s internal agent, River. The thing that makes River unusual is not its model or its benchmark score. It is that River works only in public Slack channels. No direct messages. Every conversation with the agent becomes a searchable, open transcript that any other engineer can read. As InfoWorld reported it, the scale is real. > "5,938 Shopify employees engaged with River across 4,450 Slack channels," and River "coauthors approximately one in eight merged pull requests at Shopify." InfoWorld, June 2026 Tobi Lutke, Shopify’s CEO, described the design with a line worth keeping. “The whole shop floor is the classroom.” The point is that when an agent solves a gnarly migration in a public channel, the next engineer who hits the same wall does not start from zero. They search, they find the transcript, they reuse the approach. One discovery becomes the whole org’s discovery. That is a compounding return. It does not show up anywhere in a seat-[adoption percentage](/blog/ai-coding-adoption-percentage-cto-slide-not-productivity-number). 1 in 8 merged pull requests at Shopify coauthored by an agent that works only in public, where the whole org can learn from it Now hold that against the other number that surfaced this week. The Register reported on June 25 that 97.9 percent of OpenAI’s own employees use its Codex agents, up from about 40 percent last August. For external organizations, that figure is 17.3 percent. For individuals, 0.7 percent. Same tool, and the depth of real use varies by more than a hundredfold depending on whether the surrounding environment was built to let the agent be useful. That gap is the whole point. OpenAI did not hit 97.9 percent because its people are more enthusiastic. They hit it because the company is structured so the agent’s work plugs into everything. Adoption rate measures the part that is easy to move. The hundredfold gap lives in the part nobody puts on a slide. There is a cost twist that sharpens this. The same InfoWorld piece notes that AGENTS.md context files, the small instruction files that make agents work in a real codebase, are now used by more than 60,000 open source projects, and that ETH Zurich researchers testing over 2,500 repositories found those files increased inference costs by over 20 percent. The layer that makes the agent valuable is itself a real cost and governance surface. The upside does not come free, which is exactly why it matters to know it is compounding and not evaporating. --- ## The better board number: is the agent’s work visible Here is the reframe I would offer any founder before the [next board meeting](/blog/ai-roi-where-returns-show-up-first). Stop reporting adoption rate as your headline. Report whether your agents’ work is visible enough to compound. Adoption rate answers “how many engineers turned it on.” The better question is “when one engineer figures something out with the agent, does anyone else benefit, or does it die in a private window.” A 40 percent adoption rate where the work is open and searchable will out-compound a 71 percent rate where every session happens in a silent IDE and nobody learns anything from anyone. This is not soft. It is measurable. Count how many agent sessions happen somewhere inspectable. Count whether agent-authored changes are reviewable by a named human. Count reuse, the number of times a solved problem gets found and applied again instead of re-solved from scratch. Those numbers are harder to game than a seat count, which is exactly why they are worth more. Key Insight The cheap version of adoption is "everyone has it on." The valuable version is "everyone learns from how anyone uses it." The first is a license report. The second is an org that gets smarter every week. Only one of them belongs on a board slide. ## What to change before your next board update So what actually changes on Monday. Three things, and none of them require a new tool. First, when the board asks for the adoption rate, give it, then add the second number. Say what share of agent work happens somewhere other people can see and reuse. If that number is not known yet, the honesty is itself the finding, and it points straight at the work. Second, look at where the agents run. If most sessions happen privately inside individual editors with no shared trace, the company is paying full price for the smallest version of the return. Moving even some of that work into a visible, searchable surface costs almost nothing and changes what compounds. Third, stop treating a rising adoption rate as the win condition. It is a starting condition. The win is the org learning faster than any one engineer could alone, and that only happens when the work is in the open. The founder with the 71 percent did nothing wrong by tracking it. The number is just answering a smaller question than the board thinks. The bigger question, the one that separates the companies pulling away from the ones spending money to stay flat, is whether the work the agents do ever leaves the room it was done in. That is the number I would want on the slide. #### Sources - [When software developers and AI agents share the learning](https://www.infoworld.com/article/4190178/when-software-developers-and-ai-agents-share-the-learning.html) - InfoWorld, 2026-06-29 - [Under the River](https://shopify.engineering/under-the-river) - Shopify Engineering, 2026-05-28 - [OpenAI says 97.9 percent of its employees are now using agents](https://www.theregister.com/ai-and-ml/2026/06/25/openai-says-979-percent-of-its-employees-are-now-using-agents/5262499) - The Register, 2026-06-25