Rent the pipes. Own the judgment.

The AI plumbing is going commodity, right on schedule. Managed agent platforms are here: the orchestration, the retries, the state handling, all the stuff everyone dreaded building now rents by the month. Good! Renting plumbing is usually the right call, and it was never the part that set anyone apart anyway.

But look at what does NOT come in the box. Ryan Forstie wrote a sharp piece on exactly this in August, when LangChain put a managed layer under its Deep Agents harness, and he closed on three things no platform decides for you:

"Whose domain knowledge becomes the skill, and whose doesn't."

"When to kill a subagent, a skill, or an entire workflow before sunk cost turns into political resistance."

"What 'good enough to ship' means for a given workflow."

In plainer terms: who decides how your AI does things, when do you pull the plug, and what does "done" actually mean?

I went and checked the popular agent frameworks myself, and it's starker than I expected. In most of them, "done" by default just means the model stopped asking for tools. One decides it's finished when the words "Final Answer" show up in the agent's own text. They all offer a real check...opt-in, and empty until somebody fills it. One vendor's own docs say its built-in completion checker can only judge what the agent has already put in the conversation. Credit to them for saying it out loud.

So the pipes come with the judgment seat EMPTY. That's not a knock on the pipes. It's just not their job.

And September made the same point from the very top. The people building these models spent the month telling us to slow down, shelving an IPO, and moving markets with their own words. Whatever happens up there, those three questions don't move. They're on your desk Monday morning either way.

So rent the plumbing. Just don't assume the judgment layer comes with it. Their plumbing works for everybody. Your answers to those three questions only work for you, which is exactly why they compound.

My working answers to all three, next post.

Three questions no AI platform can answer for you, and how I answer them in my own shop. Steal any of it.

Recap: the pipes rent by the month, but whose knowledge becomes the skill, when to kill a thing, and what "done" means are still sitting on your desk.

Whose knowledge: yours, landing as structure, with a human at the gate. Every lesson my AI team picks up gets PROPOSED as a change to the playbook it starts from, and one human, me, merges it or says no. The machine proposes. It never gets to approve its own curriculum. And decide what right looks like before the automation exists to flatter you.

When to kill: decide before you're invested. My experiments carry a rule, written down before the data exists, that says exactly what result ends them. My main research project has one right now: if the number doesn't move, the project closes. No arguing with sunk cost, because the argument got settled back when nobody was attached yet. Kill criteria written after launch aren't criteria. They're negotiations.

What done means: a floor and a finish line, set in advance, including the outcome where it didn't work. And here's my own miss on this one, because it's instructive: the finish line I originally picked for that project turned out to be a number that couldn't move at all, given where things stood. Took a hard look at my own measurement to catch it. If "done" is whenever it feels done, you've rebuilt the machine's worst habit...stopping at the first answer you can defend...one level up :)

And under all three: keep the record yours. Whatever you rent, the logs of what actually happened need to land somewhere you control, in a format you can read without their dashboard. Every call above gets made from that record. A history that only lives on someone else's dashboard is testimony, not a record.

That's the layer that compounds. The pipes get cheaper every quarter, and anyone can rent the same ones. Your answers get better every month you run them, and they're yours.