The Teams Most Ready for AI Agents Are the Least Likely to Let Them Run Alone

5 min Read

CNCF and SlashData released the State of Cloud Native Development Q3 2026 report at KubeCon + CloudNativeCon North America in Salt Lake City last week. The headline number is 21.2 million cloud native developers, about 41% of developers worldwide.

The finding that matters more for engineering leaders sits deeper in the report. The teams with the most mature platforms use AI far more than everyone else, and they're the most careful about handing work to autonomous agents. If you run engineering in a regulated business, that's the pattern you want to copy.

Platform maturity predicts who adopts AI at all

75% of cloud native developers use AI tools in their coding, compared with 55% of non-cloud native developers. The gap is wider for building AI into products: 67% of cloud native developers are adding AI to their services, versus 38% of everyone else.

Once teams start, though, they look almost identical. Among developers adding AI to their services, 32% of both groups have external-facing agents in production. SlashData's conclusion is that cloud native status predicts whether a team engages with AI much more than how it works with AI once it does.

We see the same thing with clients. Most of the work of adopting AI is the work of having somewhere safe to run it: standardized environments, golden paths, automated testing and deployment pipelines that already enforce your rules. Teams that have that foundation start sooner. Teams that don't spend their first AI project rebuilding it.

Most teams say they have a platform

93% of developers say their organization uses some form of DevOps standardization or internal developer platform, up from 77% a year ago. That sounds like the problem is solved.

Read the wording closely. "Some form" covers everything from a shared CI template to a fully governed self-service platform, and only one end of that range is ready for agents. We'd treat the 93% as a measure of intent. The useful question is whether your platform can give an agent an identity, limit what it can touch, log everything it does and roll back its changes. Most can't yet.

The mature teams choose supervision

The report asked developers who use AI for coding how they usually work with it. Cloud native developers were more likely to have AI draft code and then review it themselves (19% vs. 15%) and more likely to write a spec and have AI build against it (16% vs. 13%). They were less likely to set a goal and let an agent plan and complete the steps on its own (15% vs. 18%).

North America shows the pattern most strongly. 21% of cloud native developers there draft and review, and only 12% hand coding tasks to agents, one of the lowest rates in the survey.

The likely explanation is experience. Engineers who've run production systems for years treat a new, nondeterministic contributor the way they'd treat any new hire: they review its work, give it a clear spec and widen its access as it proves itself. Autonomy gets earned.

The spec-driven number deserves attention too. A written spec is what makes agentic delivery governable. It gives a human something to approve before the build starts and something to check the result against before it ships. In financial services, those two approval points are what an auditor will ask about when they want to know who signed off on a change.

Agents are already in production

Among developers adding AI to their services, 49% already have agents in production: 32% facing customers and 17% serving internal teams. Another 21% are testing. Only 6% have no plans to use agents.

Developers who haven't added AI yet point to data privacy and security first, then cost. That ordering matches what we hear from fintech and financial services teams. The question has moved past whether agents will reach production. It's whether governance got there at the same time.

A governed agent platform needs managed identity for each agent, scoped permissions, controlled access to tools, a full activity log, automated evaluation and the ability to roll back. Those are platform engineering problems, and they're the same ones your team already solves for human developers.

Today's agent pricing won't last

SlashData is blunt about cost. Model providers and hosts are subsidizing agent usage today with venture capital and hyperscaler money, and the report expects a period of optimization, efficiency and possibly retreat once teams see unsubsidized prices. It also notes that ROI is hard to measure right now, partly because it's become difficult to see where costs are concentrated. The recommendation is to invest in cost transparency and guardrails before that shift arrives.

You can't cut spend you can't attribute. If an agent run triggers a chain of model calls, tool calls and infrastructure, you need to trace which workflow started it, what it touched and what it cost. That's observability for agents, and it doubles as your audit trail. For regulated teams, it also means telemetry about agent behavior should stay in your own environment rather than going to a third-party service.

Don't wait for a modernization program to finish

The report's advice to executives is worth repeating. Don't gate AI work behind a cloud native modernization roadmap, because the work is already happening inside your organization, sanctioned or not. Do treat your platform maturity as a leading indicator of how far your AI capability can go.

We'd go one step further: run the two efforts together, and let each one set the pace for the other. Start agents on bounded, reversible work like test generation, CI failure diagnosis and code review. Build the identity, logging and rollback controls underneath them as you go. Expand autonomy where the evidence supports it, and keep it narrow where it doesn't. An organization can only be as autonomous as its weakest layer allows.

Where does your team stand?

If you want an honest read on how ready your platform, delivery process and requirements are for agentic work, take our Agentic Engineering Maturity Assessment. It takes a few minutes and shows which area is holding you back. To see how we help teams build governed agent platforms, visit our agentic engineering page.

Data in this post comes from SlashData's State of Cloud Native Development Q3 2026 report, sponsored by the Cloud Native Computing Foundation. Survey fielded July to August 2026 with more than 14,000 developers in 120 countries.

‍

TOPICS:
Blogs
BY
Dan Rye
How to Cut Observability Costs Without Losing Visibility
Dan Rye

How to Cut Observability Costs Without Losing Visibility

Observability bills grow with consumption, not value. Here's where the spend comes from and how Tensure helped Pindrop save $1M a year without losing visibility.

BY
Lesley Mackie
Three Things Your Platform Needs Before Agents Touch Production
Lesley Mackie

Three Things Your Platform Needs Before Agents Touch Production

Agents can't run in production on a platform built only for humans. Tensure's co-founders lay out the three platform capabilities that get agentic engineering out of the pilot stage.

BY
Justin Billig
Agents Are Your Newest Platform Users. Build Them a Golden Path.
Justin Billig

Agents Are Your Newest Platform Users. Build Them a Golden Path.

AI agents need a golden path too: their own identity, scoped access, audit trails, and a smaller blast radius. Here's what platform teams should build.

Smooth shipping is a few steps away

Let's see how we can help your team move faster. From developer platforms to cloud infrastructure and AI solutions that get your developers shipping again.