90% of Developers Use AI Agents. Your Engineering Org Is Still a Laggard

The classic technology adoption curve has a new chapter. AI coding agents went from novelty to near-universal individual use in about a year. But using an agent and running an engineering organization on agents are two different curves, and most companies are only halfway up the second one. Here is where the industry really sits, how people are learning to work with agents, and what it takes to cross the chasm.

A quick refresher on the adoption curve

Everett Rogers’ diffusion of innovations model splits any population into five groups by how readily they take on something new: innovators (2.5%), early adopters (13.5%), the early majority (34%), the late majority (34%) and laggards (16%). The groups on the left tolerate risk and rough edges. The groups on the right want proven, low-risk, well-supported solutions.

Geoffrey Moore later added the idea of the chasm: the gap between early adopters, who will tolerate an incomplete product for a competitive edge, and the early majority, who need something that works reliably out of the box. Many technologies stall right there.

Individual adoption already crossed the chasm

On individual use, the debate is over. JetBrains’ 2026 Developer Ecosystem Survey of more than 15,000 professional developers found that, as of May to July 2026, 90% used AI coding agents at work at least weekly and 68% used them daily. Claude Code alone went from 18% to about 39% of professional developers at work between January and mid-2026, and Codex grew roughly fivefold, from 3% to 16%.

Other signals point the same way. Ramp’s spend data shows 87% of businesses buying a coding agent now buy Claude Code. OpenRouter reported that agentic token usage passed human usage on February 6, 2026, then grew about 14x in six months.

But look at how the numbers drop as the definition gets deeper. Stack Overflow’s 2026 survey of more than 30,000 developers reportedly found 66% use coding assistants or agents, while only about 26% use automated agent workflows. Using an agent in your editor is mainstream. Running agents as an automated part of how the team delivers software is still early-adopter territory.

The tool curve and the practice curve

There are really two adoption curves in the AI SDLC. The first is tool adoption: does a developer use an agent? That curve is nearly complete. The second is practice adoption: has the organization changed how it plans, specifies, reviews, ships and measures work so agents actually improve outcomes? That curve is where most companies sit right now, somewhere around the chasm.

The gap between the two curves explains a lot of what engineering leaders are feeling. Code volume is up, but delivery, quality and review have not kept pace. Linear CEO Karri Saarinen recently wrote that his organization feels “DDoSed from the inside” by how much change people and agents can generate. GitHub, at Universe 2026, said completed work, not more code, is the output that matters. Both are describing an organization that adopted the tool faster than the practice.

What each group looks like in the AI SDLC

Here is how we see the practice curve, based on what we observe across engineering organizations. Most teams have people in several of these groups at once.

Learning agents is learning to delegate

Moving up the curve is not about learning a new tool’s buttons. It is about learning a new way of working: deciding what to delegate, explaining it clearly, and verifying what comes back. Each step up requires a different skill.

The data shows the same pattern. Coverage of the JetBrains findings notes that experienced engineers getting the most out of agents give them structured specifications, verification criteria and architectural guidelines, while less experienced developers lean toward AI-assisted rather than fully agentic workflows. The same coverage reports that 46% of developers feel a reduced sense of code ownership because of AI, a sign that the identity side of this shift is as real as the skills side.

The top step is different in kind. Fleets and automated workflows are not something one person adopts. They require shared context, standards that agents can read, review practices built for higher volume, and a way to measure whether it is all working. That is exactly where the chasm sits.

The tools are now being built by agents too

One sign the innovators are already on the next curve: they are using agents to build the tools they work with. Internal dashboards, review bots, scripts, migrations and even whole workflows now get built in an afternoon by someone directing an agent. That speeds up learning dramatically, because teams can try a new way of working the same day they think of it.

It also widens the gap between groups. Innovators compound: better agents build better tools, which make the next agent task easier. Teams still on the late-majority part of the curve fall further behind each quarter, not because they lack access, but because they lack the practice.

Five ways to move your organization up the curve

The second curve is the one that counts

The first AI adoption curve, putting agents in every developer’s hands, is nearly finished. The second, turning that into a new way of building software across the organization, has only just reached the chasm. The organizations that cross it first will not be the ones with the most licenses. They will be the ones that learned, measured and shared their way up the curve fastest.