90% of developers now use AI coding tools at work. 74% of enterprise agentic AI deployments hit positive ROI within a year. And Cursor, a tool that didn’t exist three years ago, hit $2 billion in recurring revenue.
Those aren’t vendor slides. Those are real numbers. And they’re telling us something that’s easy to miss amid all the agentic AI hype and the hand-wringing about failure rates.
Coding agents are working. Actually working. Not in demos, not in sandboxes, not in carefully staged press briefings. In production. On real codebases. With measurable results.
I’ve spent the last week digging into the data on this. And the more I look, the more it seems like software development is the one domain where agentic AI has genuinely crossed the chasm. Everything else — customer support, finance, supply chain — is still climbing the hill. But code? Code is different.
And I think I know why.
The Numbers Don’t Lie
Let me give you the raw data first. Then I’ll tell you why I think it matters.
Stack Overflow’s 2025 survey found 84% of developers use or plan to use AI tools — up from 76% the year before. Nearly half use them daily. JetBrains, in their January 2026 survey of 10,000 professional developers, put the number even higher: 90% regularly use at least one AI coding tool at work.
But adoption is table stakes. The productivity numbers are where it gets interesting.
Microsoft Research ran a controlled experiment where developers built an HTTP server in JavaScript. Those with GitHub Copilot finished 55.8% faster. Accenture and GitHub jointly studied enterprise deployments and found developers using Copilot completed tasks 56% faster on average. These are not self-reported satisfaction surveys — they’re timed, controlled measurements.
And then there are the aggregate numbers. Developers using AI coding assistants save a median of 5 to 8 hours per week. TELUS, a Canadian telecom with 57,000 team members, created over 13,000 custom AI solutions internally, accelerated engineering code delivery by 30%, and accumulated 500,000 hours in total time savings. Half a million hours.
Rakuten’s engineering team tested Claude Code on a gnarly real-world problem: implementing activation vector extraction in vLLM, a codebase that spans 12.5 million lines across multiple languages. Claude Code did it in seven hours of autonomous work. With 99.9% numerical accuracy. No human contributed code during execution. Seven hours for something that would’ve taken a team of senior engineers several days.
Time-to-market for new features dropped from 24 days to 5 days. That’s a 79% improvement. Those aren’t percentage points on a slide. Those are calendar days. Real shipping velocity.
Cursor’s growth tells the same story from a different angle. $2 billion in recurring revenue. It went from a niche editor to a platform that developers at companies with over 5,000 employees are adopting faster than any tool in recent memory. Claude Code adoption at work hit 24% in the US and Canada. Its customer satisfaction score: 91%.
And the market? The AI agents market was $7.84 billion in 2025. It’s projected to hit $52.62 billion by 2030. That’s a 46.3% CAGR. A lot of that is riding on coding agents delivering value that’s concrete enough to keep the investment flowing.
Why Code?
Here’s the thing that nobody seems to be saying out loud. But it’s obvious once you look at what’s working and what’s not.
Code is structured. It has syntax. It has tests. It compiles or it doesn’t. It passes CI or it fails. The feedback loop is measured in seconds. You don’t need a governance committee to tell you whether the agent did its job — the build pipeline tells you.
Compare that to a customer support agent that needs to navigate CRM systems built over 14 years, with deprecated fields nobody cleaned up, across departments where nobody owns the handoff. Or a supply chain agent trying to optimize across vendors that don’t expose APIs. Or a finance agent parsing invoices that come in as scanned PDFs with handwriting.
Code is the cleanest target for agentic AI. Not because the problems are easier — a 12.5-million-line codebase is not easy — but because the environment is legible. The rules are explicit. The feedback is instant.
A coding agent writes a function. Tests run. Tests pass or fail. If they fail, the agent reads the error, adjusts, tries again. This loop — act, observe, correct — is exactly what agentic AI architectures are built for. And it works in code because the observation step is deterministic and fast.
I was talking to a friend who runs a 30-person dev team, and he put it bluntly: “We tried an agent for our support pipeline. Two months of integration work and it still couldn’t handle edge cases. We tried Copilot for our engineers. It worked on day one.”
That’s the difference. Day one versus month six.
The Open Source Story
The proprietary tools get the headlines — Cursor’s $2B ARR, Copilot’s 29% adoption rate, Claude Code’s 18% and climbing. But there’s something happening in open source that’s maybe more important.
Cline, Roo Code, Aider, Continue, OpenHands — these are open source coding agents that are seeing real adoption and real contribution velocity. They’re not VC-funded unicorns. They’re community projects. And they’re getting better, fast.
Aider in particular has been around since 2023 and has refined its approach to the point where serious developers use it daily. The open source ecosystem for coding agents is healthier than for any other category of agentic AI. There’s competition, diversity of approach, and rapid iteration.
Why does this matter? Because it means the coding agent space isn’t a winner-take-all platform play. It’s more like the text editor market — multiple viable options at different price points, with different philosophies. That’s healthier for users and it accelerates the whole category.
And it means that even if the big players raise prices or change terms, the tooling won’t disappear. The open source alternatives are real and getting better.
The Enterprise Adoption Pattern
Zapier deployed AI agents across their organization and hit 97% adoption. 97%. That’s not “we made it available and some people tried it.” That’s “basically everyone uses it.”
Fountain, a workforce management platform, used a hierarchical multi-agent architecture with Claude to achieve 50% faster screening and 2x candidate conversions. That’s not code generation — that’s an agentic workflow built on the same principles that make coding agents work. Structured inputs, clear evaluation criteria, and rapid feedback.
The pattern repeats: start with something boring and measurable. Don’t try to build an agent that does everything. Build one that does exactly one thing, prove it works, expand.
Anthropic’s 2026 report on agentic coding trends identifies the shift explicitly: “engineering teams are shifting from writing code to coordinating AI agents that handle implementation.” The developer becomes an orchestrator, not a typist. That’s not a threat. It’s a force multiplier.
And here’s the number that stuck with me: 60% of developer work now involves AI, but developers only fully delegate 0-20% of tasks. The sweet spot isn’t full autonomy. It’s collaboration. The agent does the heavy lifting on implementation. The human does architecture, design review, and strategic decisions.
What This Means for Agentic AI
I keep coming back to the contrast. On one side, coding agents hitting 99.9% accuracy on 12.5 million lines, saving 500,000 hours, cutting delivery cycles by 79%. On the other, the 68-percent gap between “we’ve adopted AI agents” and “they’re actually in production” that I wrote about last time.
Why the difference?
Code is the right environment for the current generation of agentic AI. The models are good enough, the feedback loops are tight, the success criteria are clear. It’s not that the models need to get better — though they will. It’s that the problem domain matches the technology’s strengths.
Customer support agents need to navigate organizational politics. Finance agents need to handle edge cases that nobody documented. Supply chain agents need to talk to systems that were never designed for API access. These are all real constraints that aren’t going to be solved by a better language model.
But code? Code compiles. Tests pass or they don’t. The feedback is immediate and unambiguous. That’s catnip for an AI agent.
I don’t think coding agents are going to eat all of software development. But I do think they’re showing us what’s possible when you match the right problem with the right architecture. And honestly? I think the gap between coding agents and everything else in agentic AI is going to widen before it narrows.
The coding agent success story isn’t just good news for developers. It’s proof that agentic AI can deliver real value, at scale, right now. The challenge is figuring out how to get the rest of the agentic AI landscape to work in environments that aren’t as clean as a CI pipeline.
But that’s a problem for another article. For now, the coding agents are shipping. And the numbers are kind of unreal.