The Model Context Protocol crossed 97 million monthly SDK downloads this month.
That number matters more than any benchmark score or product launch. MCP is boring — the kind of infrastructure that only gets attention when something breaks. It's the thing developers reach for when they're building systems that need to run reliably, not systems they're experimenting with. 97 million monthly SDK downloads means people are building for production, not proof-of-concept.
NVIDIA GTC confirmed it from the other side. The dominant conversation wasn't which model scored highest on MMLU. It was enterprise deployments — agentic systems in production workflows at real companies, handling real work, making decisions that affect real outcomes.
The experiment is over. Agentic AI is live.
And most companies are not staffed for what that actually requires.
What the Pilot Phase Got Wrong
The pilot phase of AI adoption had a recognizable vibe: someone curious and technically comfortable, identifying use cases, standing up demos, showing leadership something impressive enough to approve the next round of investment. That person was usually a developer with good communication skills and a lot of initiative. Maybe a "Head of AI" title got invented for them.
That job is done. Or rather, it's been replaced by a different job.
Running AI in production is a fundamentally different activity than running experiments. When an agent fails in a pilot, you learn something. When an agent fails in production, customers notice, SLAs break, and someone has to explain it to the board. Jeremy Ung, CTO at BlackLine, has seen the transition up close: "Pilots are often really promising. You get exciting results in an isolated environment. Scaling is where I see most of them fail."
And when they fail at scale, the organizational response is nothing like what happened during the pilot. "At that moment, everyone looks to the CIO first to fix it, then to explain how it happened," says Chris Drumgoole, former global CTO of GE and now president of global infrastructure services at DXC Technology.
"If you treat agents as if they're just clever models, you'll lose control quickly. They have to be engineered like software systems — with defined objectives, guardrails, testing, and monitoring — because that's the only way they earn trust at enterprise scale," says Michael Finley, CTO at AnswerRocket.
Production agentic systems need observability. You have to know what the agent is actually doing. They need defined failure handling for when the agent makes the wrong call. They need a governance layer — which decisions can be autonomous, and which need a human in the loop? They need drift detection, because the model you calibrated six months ago isn't the same production environment you're running today. And they need rollback capability: a way to pull an agent out of a workflow without taking the whole system down.
These are reliability engineering problems. They're also organizational problems. And the people who know how to solve them across AI systems and enterprise operations barely exist yet as a job market cohort.
The Supply Problem
The companies that went deep on agentic AI in 2024 and early 2025 — Salesforce, ServiceNow, the handful of enterprises that didn't wait for the crowd — are the only ones with people who have real production experience. Those people are not leaving. They're the ones keeping the lights on for systems that are actually running.
Everyone else who's now deploying is hiring into a market where the supply of real production AI ops experience is effectively zero. There's a lot of supply of people who ran pilots. There's almost none who've managed agentic systems in production at any scale.
This is going to price weird. The premium for someone who's actually done this — who's built observability into an agentic pipeline, managed a governance framework for autonomous decisions, handled an agent failure in a live production environment — is going to be significant and irrational by traditional comp benchmarks. Because you're not paying for their resume. You're paying for the fact that nobody else has what they have.
What CTOs Should Actually Do Right Now
Stop looking for AI project leads. Start looking for AI reliability engineers and AI systems operators — people with a background in SRE or platform engineering who've spent the last year or two building agentic systems, not just using AI tools.
The resume signals that matter:
- Has built with MCP or similar integration protocols, and LangChain, LangGraph, or similar orchestration frameworks in production (not in a demo)
- Has written runbooks for AI system failures
- Has worked on human-in-the-loop decision systems, not just fully autonomous ones
- Has dealt with model drift or agent calibration in a live environment
If you're sourcing from job boards, you're going to miss most of these people. The ones who've actually done this work are either employed at companies that were early, or they're in circles where work finds them. Standard sourcing won't reach them.
The companies that figure out this hire in the next 60 days will be running stable, scalable agentic operations by Q3. The ones still running pilot programs in Q3 because they couldn't find the right people are going to spend a very uncomfortable 2027 watching competitors who got there first.
VC5 Consulting works with technology leaders trying to make exactly this transition — finding AI reliability engineers and systems operators when standard sourcing won't find them. If you're in that hiring moment, let's talk.