The Agentic AI Reckoning Nobody Saw Coming

Podcast: Play in new window | Download

Subscribe: Apple Podcasts | RSS

I keep having versions of the same conversation. The names and logos change, but the underlying tension doesn’t: organizations are deploying AI agents fast, they’re deploying them into production, and a lot of them weren’t ready when they did it.

Monte Carlo‘s co-founder and CTO Lior Gavish joined me on the TechSpective Podcast recently, and we got into why that’s happening and what it actually means. Monte Carlo published the Agents in Production report, and the numbers are worth paying attention to. Nearly half of enterprises surveyed already have agentic solutions running on mission-critical work — not pilots, not proofs of concept. And somewhere around three-quarters of them said they deployed before they felt ready.

That’s not a surprise, exactly. The pressure to move is real. Boards are asking about AI strategy. CEOs are mandating adoption. The competitive argument for waiting is getting harder to make. But there’s a difference between accepting that reality and assuming the governance infrastructure you need is going to materialize on its own.

Part of what makes agents different from every other enterprise tool is that they don’t follow a script. You can sandbox traditional software, test it, QA it, and have a reasonable expectation that what you tested is what you’re deploying. Agents take a natural language objective and go find a path. That path isn’t always the one you’d have chosen. Lior put it plainly — agents are optimizing for the mission, not for whatever guardrails you assumed were obvious. If they can reach data that technically sits within their access permissions, they’ll reach it. If they can route around a limitation by working through another agent, some of them will figure that out.

The other layer is that these systems are probabilistic. You can trace what went wrong after the fact, but the trace doesn’t give you control. Run the same agent on the same task tomorrow, and you might get a different path. The audit log is evidence, not a fix.

Where Lior and I spent a lot of time is the scale problem. One agent, you can watch. You can inspect every decision, every tool call, every output — same way you’d stay close to a new hire you’re still calibrating. But the organizations moving aggressively aren’t staying at one agent. They’re heading toward dozens, then hundreds, and at that point, the pilot-phase approach of eyeballing everything stops being an option.

The answer isn’t to slow down across the board. What Lior kept coming back to was reversibility — don’t hand agents tasks where a wrong decision can’t be unwound — and visibility, meaning you need enough observability to catch drift before it becomes a problem you’re explaining to someone else.

There’s an analogy from the conversation that stuck with me. You jumped in the car, hit the gas, and now you’re trying to install brakes while it’s moving. That’s a pretty accurate description of where a lot of enterprises actually are. The question isn’t whether to deploy anymore. It’s whether you can see what your agents are doing well enough to catch a problem before it becomes one you can’t walk back.

That’s what we got into. Give it a listen.