Why AI Harnesses Must Separate From Models | Taylor Savage
MTS
October 14, 2026
Factory Head of Product Taylor Savage breaks down why cloud agents and orchestration primitives are replacing local developer setups, how Factory isolates AI execution environments, and why AI harnesses must remain separate from frontier model providers. Turn ideas into software people love.
Taylor Savage
One thing that I deeply believe before I started at Factory, because you see it in the Zeitgeist all the time, was that the harnesses created by the model companies are gonna be the best for their models. So CloudCode is gonna be the best for anthropic models, obviously. And Codex is gonna be the best for open AI models, obviously. And like, starting at Factory, I realized like, that is just not true at all. You can actually do a lot better than those harnesses, certainly on specific dimensions and specific types of workloads, but also even kind of in the general case for a number of reasons.
SPEAKER_2
Hello everyone, and welcome back to MTS. Today, we are joined live in person here with Taylor Savage, who is the head of product at Factory. And Factory is building the AI coding agents called Droids. I like them because they're, it's like these, I don't know, robot little ships. And this week, they're rolling out a series of launches focused on what they call orchestration primitives, the infrastructure needed to run, manage and coordinate agents in the cloud. So welcome to MTS.
Taylor Savage
Thank you, glad to be here.
SPEAKER_2
Taylor, awesome, excited to have you. Right before we went live, we were talking about droids, we're talking about Fleet Week, and we're talking about what an awesome week that you guys get to tie it within all of these launches. Lots of big stuff happening around this term that you guys coined as orchestration primitives. How do you exactly define what exactly an orchestration primitive is?
Taylor Savage
Yeah, yeah, good question. So the overall goal of Factory from the very beginning has been help you build self-improving software. And a big part of getting that point to be able to build software for yourself is the concept of a software factory, Factory AI, kind of coining that term.
And to get to the point where you have a software factory, you have different stages of the software development life cycle, all kind of operated by your coding agents, you need a way to orchestrate and organize how those coding agents work together. So how they work on the specific task that they're given, how they get their kind of long range set of marching orders, how they verify that the upstream coding agents did the right thing to move the process downstream. So generally speaking, broadly, we think of all the factory platform orchestration primitives as all the capabilities that we give the customer, the user, the developer to be able to run these bigger sets of coding agents in parallel all at the same time, all in ways so that they're all pushing in the same direction and not running in different directions.
SPEAKER_2
Yes, especially so they're not running in different directions. This is incredibly important, especially as enterprises and developers are running fleets of agents. They're not just running one agent anymore to do one task for them. They're often running a bunch of agents in parallel. And there's all these different considerations when it comes to governing, orchestration, how they actually run. I know you're launching a ton of products around this specific problem. But I wanted to get your takes as well on why you guys are so excited about agents running in the cloud. This is particularly important for what's happening and very timely.
SPEAKER_3
So I wanted to get your takes.
Taylor Savage
Yes, yes, yes. So I think it's been a running joke in San Francisco that you see engineers walking around with their laptops.
Someone sent a little Amazon. You can buy on Amazon a little device that slides on your laptop and keeps it like an itch open. So it's not even a meme anymore. It's true. It's the reality.
We're past that now. I think generally speaking in the zeitgeist, folks have woken up to the power of cloud agents, of running these coding agents in the cloud, not on your local machine. And that lets you do a lot of things, not least being able to close your laptop, walk away, have work run overnight, have work run in parallel, have work run across many agents at a massive scale that your individual laptop might not even be able to handle. So that's what we're excited about. But also this idea of the ability to run agents in the cloud is particularly relevant to when you're building this kind of larger software factory concept. Because if you think about it, if you want self-improving software that's happening in the background without an individual engineer having to like look at a ticket, take that ticket, turn it into a prompt, wait for the response to come back, go back and forth at the agent and kind of be that meat proxy. If instead you want a lot of this to happen through these kind of assembly lines in the cloud, you need agents to be able to run, obviously, on cloud machines and not on some person's individual laptop. So that's a really core problem to the problem of building a software factory is where do you run these agents? How do you make sure they're run in a secure and a controlled way?
Read the other 28 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796258177