Everyone shipped an agent stack in March.
You still own the hard parts.

If you read the trade press in late March 2026, you'd think the enterprise agent problem was packaging. Nvidia talked about an open agent toolkit and a wall of software logos at GTC. Oracle expanded Fusion with agentic applications and a studio story aimed at people who don't live in IDEs. Accenture and Databricks formalized a push to scale multi-agent patterns with lakehouse gravity.

None of that is fake. It's real budget, real SKUs, real services revenue. Your CIO will ask about it in the Monday staff meeting.

Here's the part that doesn't make the keynote: the hard work is still the boring triangle we keep writing about. Evals. Integration to messy systems and messier documents. Ownership when the thing runs at 2 a.m. The vendor story accelerates where you might host a workflow. It doesn't dissolve whether that workflow is correct, safe, or maintainable in your data.

What actually changed in March

Silicon and ISV bundles. When a chip vendor and an ERP giant both ship "agent" narratives in the same two-week window, the signal isn't "buy this box and you're done." It's "the default integration surface for the next five years is agent-shaped." That pushes you toward standards (tool protocols, identity for non-human principals, observability on spans, not just final answers).

Services plus data platforms. When a global SI and a data vendor team up to "scale agents," they're selling capacity and patterns. Someone still has to define success metrics, map tools to real APIs, and decide what happens when the agent hits a jurisdiction-specific PDF or a half-empty CRM field. That's not cynicism. It's the job.

What we tell teams the week after the announcement

1. Separate the demo from the contract. Ask what "GA" means for your tenancy, data residency, rate limits, and support when a run fails halfway through a twelve-step flow. If the answer is a PDF, you don't have an architecture yet.

2. Put evals in the purchase path. If procurement is signing before you've defined task success and failure budgets, you're buying hope. We've watched teams recover months of drift by freezing features for two weeks and building a golden set from production-shaped inputs. Boring. Effective.

3. Name an owner with pager muscle. Platforms don't wake up for incidents. People do. If nobody is accountable for rollback, prompt version drift, and vendor quota surprises, you'll get a beautiful pilot and a fragile production story.

Why this matters for NLT-style work

We build toward products and take selective consulting when we can move the outcome. The March headlines don't change that math. They just raise the ambient noise. Our customers are better served when we help them see which parts are commodity (runtimes, starter agents, reference architectures) and which parts are still bespoke (your documents, your compliance edge cases, your org chart).

If the vendor stack gets you 60% of the way, great. We'll help you with the 40% that still determines whether anyone trusts the system six months from now.

Context from the news cycle (late March 2026): Nvidia GTC agent toolkit coverage (VentureBeat), Oracle Fusion agentic applications and Agent Studio expansion (Oracle / PR Newswire), Accenture and Databricks agent scaling group (TechDay).