Agentic AI Is Forcing the Software Stack Back Together

Futurum research cited in The Great Unification found 20% of enterprises were actively running the Model Context Protocol (MCP) in production in the first half of 2026, with another 26.9% building pilots, according to the DevOps.com article. The article argues agentic AI is dissolving traditional team boundaries because agents can cross tool and organizational seams, forcing decisions on identity, ownership and control. It notes the cloud native ecosystem is absorbing AI workload demands through Kubernetes, Dynamic Resource Allocation, Kueue and the Gateway API rather than an isolated enterprise AI stack.

TL;DR — Key Takeaways - AI is both the workload and the worker. Infrastructure teams must support models as production services while also governing agents that can act across code, telemetry, deployment and security systems. - Agentic AI is breaking down traditional team boundaries. Open protocols such as MCP allow agents to cross tool and organizational seams, forcing clearer decisions around identity, ownership and control. - Full autonomy may not be the destination. The emerging operating model increasingly favors agents doing more execution while humans retain approval and judgment at consequential points. For two decades, each useful layer of abstraction gave the technology organization another specialty. Virtualization produced virtualization teams. Cloud produced cloud architects and FinOps. DevOps tried to remove a wall and became a job title. Kubernetes https://kubernetes.io/ delivered portability, along with a new group of experts needed to operate it. That specialization was not organizational failure. It was the cost of solving real problems. Yet it left many enterprises with highly capable teams working through different tools, controls and vocabularies on adjacent parts of the same system. Agentic AI does not respect those boundaries. An agent may read an issue, inspect telemetry, change code, open a pull request and check a deployment without caring that five teams own the systems involved. Its task crosses the organization because the software can. This is why the current AI transition is more than another productivity cycle. AI is both the workload and the worker, and the stack has to support both roles at the same time. The Workload and the Worker Have Different Demands As a workload, AI needs familiar operating capabilities in unfamiliar proportions. Models require accelerator-aware scheduling, cost-sensitive routing, lifecycle controls, observability and data governance. The cloud native ecosystem is absorbing those demands through Kubernetes, Dynamic Resource Allocation, Kueue and the Gateway API rather than building an isolated enterprise AI stack. As a worker, an agent needs something more. It requires identity, permission boundaries, tool contracts, auditability and a definition of when a human must approve the next action. It can move through systems at machine speed, which means a vague policy or broad credential becomes an operational risk very quickly. Put the two together and the old team boundaries become poor control boundaries. Infrastructure, platform engineering, DevOps, software development, QA and security still have distinct expertise, but none can operate the system alone. Interoperability is Becoming Organizational Connective Tissue The Model Context Protocol https://modelcontextprotocol.io/ is worth watching for more than its adoption curve. Futurum https://futurumgroup.com/ research cited in The Great Unification https://devops.com/wp-content/uploads/2026/09/The-Great-Unification-linked.pdf found 20% of enterprises actively running MCP in production in the first half of 2026, with another 26.9% building pilots. That is remarkable movement for a protocol that did not exist before late 2024. MCP gives an agent a standardized way to use tools and retrieve context. The practical consequence is that agents can cross tool boundaries that often map directly to team boundaries. The issue tracker, source repository, observability backend and deployment system no longer need separate, bespoke integrations for every agent. A wire protocol will not fix an org chart. It can, however, make the seams visible and force decisions about identity, ownership and control that were easier to postpone when people carried work manually from one system to the next. Open interoperability also matters because the newest AI middleware categories are still unsettled. Gateways, evaluation platforms, agent control planes and agent observability products could become portable pieces of a common architecture. They could also become proprietary tollbooths above an otherwise open substrate. The outcome is being shaped now, while standards and commercial incentives are still in contention. Autonomy is Not the Only Measure That Matters The market often treats full autonomy as the finish line. Current adoption suggests a more sober path. Individual AI assistance remains the largest mode at 47.2% of organizations. Supervised agents account for 18.4%, semi-autonomous agents for 13.6% and outcome-only review for just 5.8%. Combined, 37.8% are already operating agents in some mode. That distribution does not describe a world of autonomous enterprises. It describes organizations moving into agentic work while deliberately keeping humans at checkpoints. In infrastructure workflows, 66% report using AI, but only 31% say the work is fully autonomous. Deployment decisions remain one of the least automated activities in the lifecycle. This may be the durable model: Agents perform more of the work while people retain the consequential decisions. That is still a significant shift because human effort moves away from executing each step and toward specifying intent, setting constraints and judging results. The same pattern appears across roles. Platform engineers broker capabilities. DevOps engineers define what a delivery process must prove. QA engineers create methods for evaluating nondeterministic output. Security teams govern non-human identity. Software engineers spend more time decomposing problems and verifying generated changes. The disciplines do not disappear. Their work converges around governing automation on a shared platform. A Unification Thesis Should be Possible to Disprove There are good reasons for skepticism. AI could reinforce existing silos if each team adopts a separate copilot. Specialized GPU economics could split AI infrastructure from the rest of the estate. Verification could fail to scale with generation. Proprietary middleware could recreate lock-in at the control-plane layer. Those are not footnotes. They are tests. If agentic adoption stalls, role-level use diverges sharply, a separate AI orchestration stack displaces Kubernetes or automated output overwhelms delivery quality, then the unification thesis weakens. For now, several independent movements point in the same direction. AI workloads are landing on cloud native infrastructure. Agents are crossing tool boundaries through open protocols. Platform maturity is associated with stronger AI outcomes. Specialists across the lifecycle are adopting AI at similar rates and taking on a shared responsibility for governing automated work. That is the case explored in Techstrong’s new special report, The Great Unification https://devops.com/wp-content/uploads/2026/09/The-Great-Unification-linked.pdf . The report develops the model in full, examines the evidence and names the signals that will tell us whether this is a lasting operating shift or a temporary convergence.