The most reassuring line from the earliest session I caught at PlatformCon Live Day London 2026 was also the least dramatic: “Not that much changes.” In a day full of talks insisting AI rewrites the rules of platform engineering, one early-morning session made the opposite case — the internal developer platform you have already built is the foundation the agentic era gets built on top of, not something it replaces.
Three Kinds of Paths
The clearest framework of the session split platform workflows into three categories, and it is worth adopting as shared vocabulary:
- Agent paths — work executed primarily by AI agents: PR reviews, spec drafting, triage. More standardized agent paths are reportedly on the way as the tooling matures.
- Deterministic paths — the CI pipelines you already run, and they are not going anywhere. Even in an agentic platform, the assumption that an agent could just spin up a production database on its own remains firmly off the table, especially in regulated industries.
- Hybrid paths — the combination: feed the error output from a deterministic step into the model, loop until resolution. This is what people actually mean when they talk about “loop engineering,” and it is a more precise term than the hand-wavy way “agentic workflow” usually gets used.
The core insight holds up under scrutiny: agents do not replace pipelines, they work alongside them in feedback loops. That is a grounding correction to a lot of AI hype that implies deterministic infrastructure is obsolete.
The Reference Architecture, Updated
The session’s IDP reference architecture kept the shape platform teams already recognize, with two additions:
- Developer control plane — portals, knowledge surfaces, skill management, and now agent observation as a first-class capability
- Integration and delivery pipelines — the same deterministic actions as always, now additionally augmented by agent-executed steps
- Resource plane — where your software actually runs, unchanged and still vital
- Tool security and tool observability — two genuinely new dedicated layers, built specifically for agentic workloads
Infrastructure as code stays exactly where it is. Some tools need to become agent-consumable — exposing capabilities in a way an agent can discover and call — but a platform team that has already built a solid IaC foundation is most of the way there already.
The Governance Plane
A companion session from the same morning sharpened the operational reality of running agents at scale. The “operating system” of a single agent breaks down into execution, context, capability, and evaluation — and hot-loading the right context while exposing the right capabilities is, in the speaker’s words, “going to be our bread and butter for years to come.”
The scaling problem is the one platform teams will recognize immediately: running one agent is a demo, running dozens or hundreds or thousands simultaneously demands a governance plane covering identity, security, and observability — the same shared infrastructure platform engineers already build for human-driven workloads, extended to cover agents before the sprawl gets ahead of you. This is the same argument I make about production guardrails for AI agents: the governance layer has to exist before scale forces it into existence reactively.
A Concrete Walkthrough: Agentic PR Review
The most useful part of the morning was a practitioner stepping through exactly how an agentic PR-review workflow runs inside a real platform, end to end:
- A human triggers the PR review, firing a webhook that starts an orchestration function.
- Identity is resolved, and the correct execution path is selected.
- Policies are pulled and available capabilities are identified.
- Context is gathered and handed off to the model for execution.
- Guardrails are applied and an evaluation runs.
- The platform decides: loop again, or report back to the human.
The standout observation — echoing Anthropic’s own writing on agentic systems — is that model execution is only a small slice of what the platform actually does. The real engineering work lives in orchestration, identity resolution, policy enforcement, and context assembly. That is precisely the layer building AI agents as Kubernetes operators is designed to formalize: treat the orchestration logic as infrastructure, not as glue code bolted onto a chat completion call.
Why This Framing Matters
Every vendor pitch at PlatformCon London wanted to convince you that agentic AI demands a wholesale platform rebuild. This session’s quieter, more useful claim is the opposite: the discipline of platform engineering — deterministic pipelines, policy enforcement, identity, observability — is exactly what makes agentic adoption safe. The work is extension, not replacement.
Related Reading
- PlatformCon London 2026: The AI Era Runs on Platforms
- Production Guardrails for AI Agents
- Building AI Agents as Kubernetes Operators
- CNCF Platform Engineering Technology Radar 2026
About the Author
I am Luca Berton, AI and Cloud Advisor. I work at the intersection of platform engineering, cloud security, and enterprise AI deployments. Book a consultation.
