The Autonomous Enterprise Meets the 30GB Memory Leak

SAP and OpenAI are pitching agents that run your business while practitioners discover Claude Code can quietly eat 30GB of RAM. The operational reality of agentic AI is catching up with the vendor narrative.

This week's agentic AI signals split cleanly into two stories that should be read together: enterprise vendors are racing to sell agents as core business infrastructure, while developers running those systems are surfacing the operational debt nobody talked about during the demo. Both stories are real, but only one of them is being marketed.

On the enterprise side, the pitch is escalating. SAP unveiled its "Autonomous Enterprise" model this week, positioning AI agents as the successor to the ERP architecture it has built around for 50 years. The same week, OpenAI launched a dedicated deployment company targeting embedded enterprise AI architectures tied to operational systems — explicitly an answer to clean-core modernization. Twilio joined the chorus at SIGNAL 2026 with a unified agentic AI platform strategy covering customer-engagement infrastructure.

Then there is the practitioner view. Simon Willison reported that his Mac was running short on memory because idle Claude Code processes across his terminal windows were consuming ~30GB combined — the largest single instance at 4.9GB, with default settings and no agent features enabled (as of May 2026). Meanwhile Jason Liu added a pattern to his agents.md instructing his agent to "start a loop and keep an eye on" a Slack thread until someone replies — a quietly clever workflow that also means an LLM is now polling indefinitely on his behalf. These are not edge cases. They are what running agents actually looks like today.

The gap matters because of what TechRadar calls "objective drift": semi-autonomous systems deviating from their intended goals when running unattended. Many enterprise services, the same piece notes, aren't yet "agent-addressable" — meaning the autonomous-enterprise pitch is selling a destination the underlying APIs and governance frameworks haven't reached. A separate practitioner framework, Arnav's Securing Agentic AI guide, describes these systems as initiating activity through "multiple autonomous loops with persistent memory, tool chaining, and multi-agent orchestration" — each of which expands the attack surface and the operational footprint at the same time.

For builders, the interesting work is at the seams: monitoring loops, memory budgets, drift detection. For buyers evaluating the SAP or OpenAI enterprise narratives, the right question isn't whether agents can run a process — it's whether your team has the audit, observability, and rollback story for when they don't.

The bottom line: Treat agentic AI like database infrastructure, not magic — budget for memory, monitoring, and human oversight before you let it touch the books.

Sources