We're rebuilding the DevOps mess, one agent at a time
Autonomy for every team means paralysis for the enterprise
Something odd is happening in enterprise engineering. Individually, everyone is faster than they have ever been. Ask any developer and you get the same answer: AI has changed how they work, and they are not going back. Yet the big, integrated, cross-team work seems to be getting harder, not easier.
Walk any enterprise engineering floor right now and you will find everyone busily building out their AI stack. One team has wired up its own agentic coding setup with a handful of custom agents. Another has built an orchestration layer on a different model, with different guardrails. A third is rewriting its original harness to use agentic loops or maybe graph-based execution.
Every one of these setups uniquely benefits those who built it. Each is bespoke. All of them were developed in record time, thanks to AI. Leaders are ecstatic to see new tools popping up each week, brimming with pride at how deeply AI is being adopted.
I’ve sat with many teams this year, listening as they excitedly tell me about their journey. The energy is incredible, and the productivity gains are real, at least for the team standing in front of me. But I have that nagging feeling that I’ve seen this movie before, and I know how it ends…
I’ve seen this movie before…
We are stumbling into the same problem we spent two decades digging our way out of with DevOps. Different teams, different tools, different ways of working, each defended with the same reasonable line: it works for us. The difference this time is the speed. It took the industry the best part of a decade to build the infrastructure sprawl that platform engineering was later invented to untangle. The agentic version is compounding in months, because the cost of going your own way has dropped to a credit card and an API key.
More Divergence, Faster
Every developer now carries a private collection of prompts that work. Every team has its own way of coaxing an agent through a task. Almost none of it is written down anywhere or designed to be reusable. Multiply that across a large organisation and you get a peculiar kind of waste: thousands of small, clever solutions, each one solved exactly once, none of them compounding.
Cost is the most visible outcome. When every team is working with a different vendor, there are no economies of scale, nobody has buying power, and finance only sees the true total at year end.
The invisible parts are much worse.
Divergence of tooling and processes is a structural barrier to collaboration, knowledge sharing, and economies of scale.
Divergence creates and hardens organisational silos. When each team builds and maintains its own solution and has its own definition of “done”, the boundaries between teams stop being just lines on an org chart and are instead walls in the architecture. Knowledge can’t flow: a practice proven in one team can’t be adopted by another without being rebuilt from scratch. Neither can trust: one team’s “compliant” isn’t the next team’s “compliant.” So the organisation solves the same problems again and again.
Leaders will claim it’s “autonomy” when what they’ve actually bought is the inability to ever act as one.
Where To Begin
The reflex is to consolidate the tools. Pick a winner, mandate it, shut the rest down. If we learned anything from platform engineering and golden paths, it’s that getting teams to consolidate and collaborate is more than a technology problem.
In a recent article, I encouraged teams to define a set of ground rules before trying to enforce common tooling. I suggested a set of best practices that all teams should agree to, irrespective of tool choice. In AI terms, an organisation must have control of which models are being used, and what data AI tools should be permitted access to.
Model selection should not be an engineering preference. It’s a data protection issue. An amazing agent running on an ungoverned model with access to sensitive corporate data is a liability, not an asset.
We also have to be honest about how people learn a new technology. Nobody masters this by reading a policy. They master it by trying things. Tools and prompts will stay diverse for a while, and that is fine. Exploration is the price of competence.
What leaders can steer, gently, is the shape of that exploration. Encourage teams to favour tools that keep their options open. Claude Code is a good example of the tension. It is excellent, and it runs on Anthropic’s models and nothing else. OpenCode gives developers a very similar experience but sits on top of whatever model you point it at, including open-weight models you host yourself. Same craft, fewer locked doors.
That distinction matters more than it looks. A model-agnostic harness is what turns “which models are we allowed to use” from a wish into a switch. You cannot govern the substrate if changing it means re-tooling every team. Agnostic surfaces are what make a governed floor enforceable.
Treat Reuse As Capital
Leaders should stop thinking about AI as a productivity tool that everyone must master and start thinking about it as compounding capital. An enterprise’s real advantage is never a single clever prompt or a single sharp agent. It is the ability to take what one engineer learns and turn it into something a thousand engineers can build on.
Most organisations I work with have set up some sort of “library” for people to share their prompts, but in practice they seem to be of limited use.
There is a deeper reason those libraries gather dust, and it is not a lack of discipline. AI output is non-deterministic. Change the model, the harness, the memory, or the surrounding context, and the same prompt can produce something meaningfully different. A snippet of code either compiles or it does not. A prompt someone hands you is a bet on conditions you may no longer share. So people do the rational thing. They regenerate rather than reuse.
Reuse has quietly been replaced by regeneration.
Skills were meant to be the fix, the packaging format that finally lets a capability travel. As a standard, they are promising. As an enterprise asset, they are immature. Public registries accept almost anything, with no mandatory review, and it shows. Security audits routinely find flaws in a quarter or more of the skills they scan, and 2026 has already brought coordinated campaigns that seeded hundreds of outright malicious skills into public marketplaces. Worse for reuse, there is still no reliable signal for quality. When researchers found malicious skills sitting in a public registry, ordinary users had no way to tell them from the safe ones for months.
Put those together and the incentive is obvious. Unpicking an inherited skill you cannot fully trust is often slower than writing a fresh one, so everything starts to feel disposable. That is comfortable, and it is quietly corrosive, because an organisation that regenerates everything never accumulates anything.
Making reuse compound is an economic decision to be resourced.
Someone has to own the shared library. Someone has to prune it, because a hoard nobody curates rots into a junk drawer and sends people straight back to building their own. Reuse left to goodwill does not happen. Reuse that is owned, funded and measured does.
Move Now, While The Divergence Is Cheap
Divergence compounds, and so does the cost of reversing it. Every month that teams keep building in different directions adds more bespoke setups to reconcile, more undocumented dependencies, more people with muscle memory in something you will later want to retire. We let DevOps sprawl mature before we addressed it, and that delay is most of why the cleanup took twenty years. The cheapest moment to act is always the one with the least accumulated divergence, and that moment is now.
Acting now does not mean a big-bang consolidation, and it does not mean banning experimentation. Teams can still move fast, still experiment, and still build agents that suit them. The difference is that they can build on top of solid foundations, their learning lands somewhere the next team can find it, and their spend shows up in one place instead of forty. The experimentation compounds instead of scattering.
That reality is available to any organisation willing to make three decisions while the divergence is still shallow. Standardise the substrate. Resource reuse as capital. Act before the bill grows. The one thing that will not work is waiting for it to settle on its own, because we know sprawl will never resolve itself.