The Autonomous Enterprise Arrives Before Its Operating Manual

SAP just declared the end of the ERP era and OpenAI launched a deployment company. The control planes and orchestration frameworks to actually run agent fleets are still being invented in real time.

On May 12, SAP declared the end of the ERP era — its Autonomous Enterprise model puts AI agents in charge of core business operations, rewriting the 50-year ERP playbook that defined the company. The pitch landed the same week OpenAI launched a deployment company targeting enterprise operations and OpenHands rolled out an Agent Control Plane. Three vendors, one message: agents are about to run your business. The infrastructure to make that work is still being invented in real time.

The pitch versus the orchestration reality

SAP's Forbes framing is the loudest version of a trend building since late 2025: agents move from copilots to operators. OpenAI's deployment company — a services arm dedicated to embedding AI into existing ERP systems — is the practical bet that enterprises won't figure this out alone. Both stories are stronger for being announced together, but neither addresses the part developers are wrestling with right now.

That part is orchestration. A dev.to teardown of LangChain vs LangGraph calls it the "Stateless Wall" — simple chains break the moment a workflow needs memory, branching, or recovery, and production teams are migrating to stateful orchestration frameworks. Builders running nested agents are improvising the rest: Jason Liu described a setup where his main agent has a /goal and spawns 3 sub-agents each with their own. That works for one developer with the full mental model. It does not scale to 200 employees in the same Slack workspace.

Governance is the real gap

TechRadar's overview of agentic AI in 2026 flags "objective drift" — semi-autonomous systems deviating from their intended goals — as the dominant risk, and notes most current services are not yet "agent-addressable." Shopify's approach, highlighted by Simon Willison, is more interesting than the vendor announcements: their River agent system lives in Slack and can only be used in public, so other employees learn from watching. Willison compares it to Midjourney's Discord-only launch teaching people image prompting by example. That is governance as a learning channel, not a compliance layer.

Even the tools themselves are not free. Willison also flagged that Claude Code processes on his Mac were consuming ~30GB of RAM across terminal windows, with the largest single instance at 4.9GB. The agent era's resource bill is showing up on developer laptops before it shows up in cloud invoices.

The bottom line: Treat this week's enterprise-agent announcements as a signal that vendors are ready to sell the autonomous enterprise — and as a reminder that orchestration, governance, and even local resource costs are still yours to figure out before signing anything.

Sources