Agent Orchestration 2026-07-04

Q&A: What is agentic AI today, and what do we want it to be? | MIT News | Massachusetts Institute of Technology

MIT's Phillip Isola explains how AI agents actually learn — by trial and error on live systems. For GTM teams being sold "autonomous agents," this is the part the vendor deck skips.

Source: Q&A: What is agentic AI today, and what do we want it to be? | MIT News | Massachusetts Institute of Technology

The news

MIT computer scientist Phillip Isola sat down with MIT News to cut through the agentic AI hype and explain what these systems actually are and how they learn. His core point: agents don't come pre-loaded with procedural knowledge — they figure out what to do by trying things, failing, and adjusting. There's no instruction manual. The agent has to discover the process itself.

Our take

This is the part that gets glossed over in every agentic AI pitch deck aimed at GTM teams.

When Isola describes an agent learning to book a flight — visiting airline websites, clicking around, figuring out what works — he's describing a system that has to infer the process because no one handed it a documented one. That's fine for a research environment. It's a liability in a live CRM or a marketing automation stack where a wrong click can corrupt data, trigger a send, or blow a segment.

The implication for marketing and RevOps teams isn't subtle: an AI agent is only as reliable as the process it's operating in. If your lead routing logic lives in someone's head, if your qualification criteria shift based on who you ask, if your sequences were built by three different contractors over two years — an agent deployed into that environment isn't going to learn the right thing. It's going to learn whatever the messy reality of your stack reflects back at it.

This is why AI in GTM is an operator problem, not a technology problem. The research community is working hard on making agents more capable. But capability without process documentation is just faster chaos. The teams that will get real mileage from agentic AI aren't the ones who adopt earliest — they're the ones who did the boring work first: mapping their workflows, writing down the rules, and building systems that a human (let alone an agent) could actually follow.

The trial-and-error learning Isola describes is interesting science. In a production GTM environment, it's a support ticket waiting to happen.

So now what?

Before you evaluate any agentic AI tool for a GTM workflow, run this quick audit:

The agent technology is advancing fast. Your process documentation is what determines whether that speed works for you or against you.

Want to build this capability for your team?

If you want automations like this running inside your GTM stack — not just a template but a working system — book a call and we'll scope it together.

Book a Discovery Call