If your team has been working with coding agents for months, it probably deploys more. The uncomfortable question is whether it also fails less.
This piece argues a thesis that is easy to state and hard to apply: speed comes from the tool; stability comes from the method. If the team does not change how it specifies, how it checks and who decides, AI only accelerates what it was already doing, failures included.
What DORA has seen two years in a row
DORA, Google's research programme on software teams, measures delivery performance every year with two families of indicators: throughput (how many changes reach production and how fast) and stability (how many fail and how costly it is to recover). In its 2024 report it estimated that, as AI adoption increased, both went down:
DORA explained it at the time: improving the development process does not automatically improve software delivery, at least not without sticking to the basics, such as small batch sizes and robust testing mechanisms.
In 2025, with responses from nearly 5,000 professionals, one of the two numbers changed sign. DORA now observes a positive relationship between AI adoption and delivery throughput. The other did not: AI adoption still has a negative relationship with delivery stability. In other words, teams now go faster, but they do not fail less.
Risk These are correlations in surveys, not proven causes, and the 2024 figures are estimates. The 2025 announcement gives no percentage for stability. None of this says AI breaks your software. It says stability does not rise by itself when speed does.
Why speed arrives on its own and stability does not
With an agent, generating a change costs far less than before. Checking it costs the same. If nothing else changes, the team produces more changes with the same capacity to filter them, and the difference ends up in production.
DORA sums it up in its 2025 announcement: AI accelerates software development, but that acceleration can expose weaknesses downstream. And it names them: without robust control systems, such as strong automated testing, mature version control practices and fast feedback loops, an increase in change volume leads to instability. It is the same idea as its 2025 report, which describes AI as an amplifier of the strengths and weaknesses an organisation already has, and which is where we started in before the agent, the safety net.
The tool decides how many changes you can produce. The method decides how many of those changes deserve to reach production.
What changes in practice
There is no need to slow the agent down. Every change it produces has to go through a net that used to be optional and no longer is. There are four practices and they work together:
- Acceptance criteria written before the code. If what the change has to meet only lives in someone's head, nothing can check it. It is the basis of Spec-Driven Development (SDD), where the specification is the source of truth and is approved before any code is written.
- Small changes. A small change gets reviewed, tested and undone. A big one gets approved out of fatigue.
- Undo without drama. If going back is easy, a failure is a short incident. If it is not, it is a lost week.
- Automated checks before human ones. Continuous integration filters what is mechanical; the person reviews what carries risk.
The first is the one most often skipped. Writing the criteria first lets you turn them into the merge gate, and it is what we explain in Spec-Driven Development with AI. It applies equally to security requirements, as we explain in the agent's code passes the tests and isn't secure.
The second and third are the ones DORA highlights in its capabilities model for AI-assisted teams: working in small batches is especially powerful in an AI-assisted environment, and as AI increases the velocity of changes, version control becomes a critical safety net. Its recommendation is for teams to become proficient in rollback and revert features. How to prepare for that is in how to undo what a coding agent does.
The fourth changes the role of review. In an InfoQ webinar reported on 16 October 2026 it was said that line-by-line manual code review cannot keep pace with what agents produce: style and syntax checks have to move into continuous integration, and the person has to review architecture, business intent and non-functional requirements. How it fits with the rest of the process is in reviewing AI-generated code. And a person decides where there is risk.
How to tell whether this is happening to your team
The trap is measuring only half. If the team looks at how many deployments it makes, AI always wins. If it also looks at how many of those changes have to be fixed or rolled back later, the other half appears. Measure both at the same time, at the same frequency, and compare them with what you had before the agents.
Which indicators to choose is in KPIs for AI development teams. And if the team's AI spend has to be justified in the budget, stability is part of the value thesis worth asking the team for: delivering faster what then has to be redone is not a saving.
A second look
If you want to review with someone from outside what each of your team's changes passes through today before reaching production, the diagnostic is free, and if the diagnostic shows no difference, there is no Quick-start. What the rest of the path looks like is in AI-Accelerated Development.
Frequently asked questions
Does AI make software less stable?
There is no evidence that AI, on its own, breaks software. What DORA has observed two years in a row is an association: in organisations with more AI adoption, delivery stability is lower. It is a correlation in surveys, not a proven cause. The useful reading is a different one: stability does not rise by itself when speed does.
What exactly does DORA say about AI and delivery speed?
In 2024, DORA estimated that, as AI adoption increased, delivery throughput fell by 1.5% and delivery stability by 7.2%. In 2025, with responses from nearly 5,000 professionals, the relationship with delivery throughput is now positive, but the relationship with stability is still negative.
Why does speed go up and stability does not?
Because generating changes has become cheap and checking them has not. DORA puts it this way: without robust control systems, such as strong automated testing, mature version control and fast feedback loops, more change volume leads to more instability. The tool accelerates; the net that filters each change has to come from the team.
Which practices protect stability when a team uses coding agents?
Four, and they work together: acceptance criteria written before the code, small changes, the ability to undo any change without drama, and an automated check on every change before human review. Where the risk is high, a person decides.
Do we still need to review code by hand?
Yes, but not line by line. In an InfoQ webinar reported on 16 October 2026 it was said that line-by-line manual review cannot keep up with agents: style and syntax checks go into continuous integration, and the person reviews architecture, intent and non-functional requirements.
How do I know whether this is happening to my team?
Measure, at the same time, how much you deliver and how much you have to fix or roll back afterwards. If deployments go up and failures after a change do not go down, AI is also accelerating what was already failing. Looking at only one of the two numbers hides the problem.
Sources
- Nathen Harvey and Derek DeBellis (DORA), "Announcing the 2024 DORA report", Google Cloud Blog, 23 October 2024.
- Nathen Harvey and Derek DeBellis (DORA), "Announcing the 2025 DORA Report: State of AI-Assisted Software Development", Google Cloud Blog, 23 September 2025 (nearly 5,000 professionals surveyed).
- DORA, "State of AI-assisted Software Development 2025", 2025 (no exact date visible; accessed 21 October 2026).
- Kevin M. Storer and Derek DeBellis, "Introducing the DORA AI Capabilities Model", Google Cloud Blog, 23 September 2025.
- Olimpiu Pop, "AI in Production: Panellists Highlight Silent Failures, Harness Engineering, and Agent Governance", InfoQ, 16 October 2026.
- The list of four practices and the way of measuring are onext's own work.
Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, a person who decides where there is risk and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →