The era of enterprise AI agents has arrived — but not in the way the hype suggested. At the AI Agent Conference in New York this week, leaders from Datadog, T-Mobile, CrewAI, RingCentral, and others described a shift that's quietly changed the engineering conversation: building agents is the easy part now. Trusting them in production is the real problem.
"One of the hardest things for humans to do is no longer building production systems. It's actually reviewing the vibe-coded software that gets shipped into production."
— Ameet Talwalkar, Chief Scientist, Datadog
What actually changed
- The bottleneck moved upstream. A year ago, the hard thing was getting agents to do anything useful. Now the hard thing is validating what they produce before it hits production — especially as AI coding agents generate code at a pace humans can't review comfortably.
- T-Mobile runs 200,000 AI-handled customer conversations per day. That's not a pilot. That took a year to build and involves serious governance investment.
- Simulation is the new testing. ArklexAI launched ArkSim specifically to simulate AI-agent interactions before production deployment. Why? Because "agentic interactions are not deterministic" — the same agent, different customers, unpredictable outcomes.
- The framework phase is maturing. CrewAI's Joe Moura: "Initially, it was all about building and deploying agents. But now it's all about security and enterprise adoption." Agent frameworks are trending toward commoditisation; the differentiators are now enterprise features, not core agent logic.
- Hallucinations remain the enemy. Akamai CTO Bobby Blumofe highlighted that LLMs sampling probabilistically will give different answers at different times. Web-search-augmented context and knowledge graphs (e.g. LanceDB's new Lance Graph project) are increasingly used to ground agent outputs.
The real lesson from the conference
Human oversight isn't a temporary safeguard to be engineered away — it's a core design principle. Almost no speaker framed autonomy as an immediate goal; instead, they framed it as a future destination you reach by being very careful right now.
RingCentral put it plainly: "Our goal isn't to eliminate a live agent. We're trying to make their lives easier. If we can offload fifty or sixty percent of the tedious stuff, that leaves them more time for strategic work."
That's not a hedge. That's the actual product philosophy driving enterprise adoption in 2026.
The Bill Gates "AI agents = autonomy" framing from 2023 is quietly being replaced by something more practical: agents as supervised workers that need simulation, validation, and continuous human review before you'd trust them at scale.
What to do
- Shipping agents internally? Build review and simulation checkpoints before production — not after. ArkSim-style pre-deployment validation is going to become table stakes.
- Running LLM-backed agents? Treat hallucinations as a systems problem, not a model problem. Knowledge graphs and RAG are your levers.
- Evaluating agent frameworks? CrewAI, LangChain, and others are converging on enterprise features. Pick the one that fits your security and governance requirements, not just the one with the coolest demos.
- Managing AI coding agents? Talwalkar's point stings but it's right — your team's code review process is now your most important quality gate. Invest there.
✏️ Drafted with KewBot (AI), edited and approved by Drew.