Deutsche Telekom's LMOS: what enterprise AI teams actually want from a platform

Arun Joseph at InfoQ describes a real, in-production agentic compute layer: ephemeral agents, an ADL, and a central capabilities catalog.

In plain terms. Arun Joseph, presenting at InfoQ, described Deutsche Telekom's enterprise AI platform (LMOS): a layer that replaces "tool sprawl" with a set of core abstractions, and introduces an Agent Definition Language (ADL) with ephemeral agents. The framing: agentic compute is the missing layer that no cloud provider ships out of the box.

The fact

On August 3, 2026, InfoQ published Arun Joseph’s presentation on LMOS at Deutsche Telekom. Three theses: (1) companies face unchecked proliferation of LLM/agent tools ("tool sprawl"); (2) what they need is an "agentic compute" layer with shared abstractions; (3) LMOS offers an Agent Definition Language (ADL) and ephemeral agents instead of persistent services.

Our take

Joseph’s thesis isn’t new—QCon AI Boston (post #1211) and Cloudflare Agents Week (post #1766) tell similar stories—but its value lies in coming from a telecom operator that must maintain decades of legacy business code. LMOS is a production experience report, not a vendor keynote.

What this pattern means for your architecture

• Ephemeral over persistent. An agent has no inter-session state; state lives in the data plane (vector DB, business DB, event log). This avoids the pitfall of "long-running conversations leaking context" and eases horizontal scaling. The cognitive cost: treat every agent call as a pure function call, not a living object.
• ADL as a contract. Describe an agent by what it consumes and produces, not by the prompt you pass it. Two teams can then write interoperable agents. This is the real promise of MCP applied internally, with stronger typing.
• Core platform vs. tool sprawl. Instead of each team plugging tools via LangChain on one side and LlamaIndex on the other, a shared catalog of capabilities with shared observability and policies. This is platform work, not application work.

Under the hood

LMOS isn’t public yet. The closest open-source parallel is emerging around MCP + registries (Anthropic MCP hub, Cloudflare AI Gateway, Portkey). The structural difference: Deutsche Telekom controls the data plane (telephony, billing, internal CRM)—a use case a U.S. cloud can’t outsource, making internalization of the platform structural.

So what

For architects sorting between "agentic framework of the month" and a real platform: the test is this. If your agent layer doesn’t let you define an agent without writing Python code and deploy it without redeploying a service, it’s not a platform—it’s an SDK. LMOS suggests real production platforms look more like Airflow than LangChain.

To watch

Possible open-sourcing of LMOS or a component (ADL); publication of operational metrics (number of agents in production, latency, cost per task); adoption by other European operators under DMA/data-sovereignty constraints.