Wordware founders, Filip Kozera and Robert Chandler - non-engineers can build AI workflows
Wordware's Filip Kozera on why he wants an AI agent to be wrong 20% of the time instead of chasing deterministic certainty, the "potion maker vs wizard" way he explains prompt chains to non-technical users, and why he thinks Replit's biggest problem is having to fight on two fronts at once.
Filip Kozera and Robert Chandler are the cofounders of Wordware, which lets you describe a repeatable task in a document-style interface and turn it into a background agent that runs on a schedule or a trigger, then asks you to approve or reject what it did.
We get into what actually separates Wordware from ChatGPT's old plugin store, why treating AI as a fuzzy, stochastic system instead of expecting deterministic software is the real mindset shift, who Wordware is actually built for in the crowded space between Cursor and Lovable, and what it takes to be a good "Wordware aficionado" even without knowing how to code.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
What does Wordware actually do?
We're building agents that automate repeatable processes in your life. You describe the whole assignment for an agent in a document-like format, which feels friendly and familiar rather than like writing code, then you schedule it based on explicit or implicit triggers and it runs. You get a notification and can approve or reject what it did. We're trying to bring about a world where you don't do busy work, you do the creative, deep work you actually love — one of the metrics we half-jokingly track internally is the poop index: productivity optimization on potty, how much you can get done in the ten minutes you're sitting on the toilet.
How is this different from something like ChatGPT's old plugin store, which never really took off?
We used to explain it to non-technical people this way: with ChatGPT, you're a wizard casting a spell, muttering the incantation — you have to be a learned wizard who knows the right prompt engineering, and then you wait for it to happen, back and forth, every time. With Wordware, you become a potion maker: you distill the process once, mix up your cauldron, then you can bottle that potion and pour it onto new data whenever you need it, or hand the bottle to a friend so they get the benefit without having to learn the incantation themselves. ChatGPT or Claude is the tool for one-off, in-the-moment research; we're for the repeatable tasks — the ones where you're willing to put in a little more upfront effort because you know you're creating an assignment for an AI agent you'll reuse again and again, and it feels like you're training your own AI to eliminate a piece of your work rather than just complete it once.
You've built background agents for yourselves already. What does that actually look like day to day?
The first agent I wrote for myself in Wordware was based on a fairly complicated prompt that understands my priorities at this particular point in the company. Anything it scores above 90 out of 100 gets surfaced to me immediately as a Slack message with a link straight to it, so I can respond right away or from my phone. It takes away the background anxiety of wondering if there's something urgent I'm missing. A lot of what people get from hiring a human assistant — not having that constant low-grade anxiety about your inbox, being able to do deep work without worrying you're missing something important, having someone flag you only when it actually matters — can be introduced very quickly with agents like this. I recently hired my first actual assistant too, and it changed how good I am at my job; we want to give everyone that same shift, through a document interface where you explain and control the work, and a mobile app where you approve it, so you always feel in control.
If someone doesn't need to know JavaScript syntax to use Wordware, what actually makes someone good at it?
You need to be able to speak clearly to an LLM and break a problem down into its steps — if you can't communicate clearly enough to explain a task to an intern, you're probably not going to be able to automate it with AI, let alone with Wordware. We have clients who ask for things that are basically wishful thinking — one guy wanted to optimize gas consumption for every house in Poland based on the weather and just expected the AI to figure it out. If you don't know how to get there yourself, AI isn't magic, it won't get you there either. What AI is genuinely great at is taking the kind of white-glove treatment you'd normally reserve for your biggest customers — because they're worth millions in revenue — and making that available to everyone. But the knowing how is still the important part. We always tell customers: think of it like a piece of paper — write down, step by step, what needs to happen and what decisions need to be made. If you can do that, Wordware can do it. If you can't, it's just wishful thinking.
How do you think about the difference between what an ML system can do and what an LLM can do?
People mistake ML problems, which need vast amounts of labeled data, for LLM problems. LLMs are good at taking huge amounts of human-produced data, especially text, and distilling it into a summary or new content — things that can be judged good or bad, but can't really be judged right or wrong. What we're trying to do is bring enough structure to LLMs that they can make real decisions, and that's hard, because software has always been an intrinsically deterministic system you can verify, while dealing with words and humans is a fuzzy, stochastic system. You have to stop expecting everything a computer does to be deterministic, and instead weigh the benefits of a stochastic system against its downsides — every human is a stochastic system too, and we still hire people, pay them well, and they perform. AI entering that world means accepting you won't always fully understand what it did, as long as you're getting real benefit and real money out of it.
How do you actually apply that "embrace the stochastic system" mindset in practice?
We don't run rigorous evals on humans either — there are cultural rules like "80% of your work should be right, 20% you should be taking risks and being yourself." Imagine applying that same rule to agents: 20% of a task being wrong is fine, as long as the system is learning as it goes. We're getting there even with current models — with something like Codex, I'll now fire off almost any bug I see to it before I even write a ticket for the team, because there's maybe a 30% chance it one-shots it and it costs me almost no time or effort to find out. If it doesn't work after one follow-up comment, I still file the ticket, but I've lost nothing by trying first, and that hit rate only goes up from here. It really does lower your bar for a first attempt, since you're not merging anything to main without review anyway. The cost of that first attempt is also incredibly low compared to a person's time — that's what we're trying to do for knowledge work generally: run a bunch of background agents on things you've described once, and just approve or reject the result, the same way you'd gradually trust an assistant to send more of your emails without you reviewing every one first.
Who is Wordware actually built for — is this sold to engineers, or to non-technical domain experts?
We're really built for cross-functional AI product teams. What we see over and over is an engineer writing the prompts and the codebase, showing the output to a PM or domain expert, and getting back "the vibes are off" as feedback — which drives the engineer crazy, because they want concrete requirements, but the domain expert genuinely can't always articulate why one output is better than another. It's the same as an SDR who can tell a good cold email from a bad one but can't always explain exactly why one feels more human and the other feels salesy — that's tacit knowledge built over years of experience, and it shows up the same way in healthcare, legal, finance, and marketing. What Wordware does is put the domain expert in the driving seat, letting them directly iterate on the prompt chains and bake their taste into the agent, while still collaborating with an engineer who turns the result into an API that plugs into the rest of the product.
How do you think about your position relative to something like Lovable, Replit, or Cursor?
If I had to choose between being positioned like Cursor or like Lovable versus being positioned like Replit, I'd choose either of the first two — Replit is fighting on two fronts at once, needing to serve both engineers and non-technical builders, and that go-to-market is genuinely hard. We've realized we have a similar challenge, sitting in an infrastructure-ish position where we need both engineers and domain experts. So we're building a much simpler entry point where you just write what you want the way you'd write a description for an intern or a ticket for an engineer — that's rolling out to all our users soon. Wordware today still requires a bit of a systems mindset: knowing when to split a workflow into sub-flows, when to use an if-else or a loop. This new entry point leans on the models having gotten good enough in the last six months that a copilot can help draft that assignment for you, asking clarifying questions the way a research agent does, so someone without any of that systems background can still build something equally powerful.
Any closing advice for founders?
Build with friends. This isn't going to be a two-year journey with a $50 million acquisition waiting at the end — you have to build something you genuinely believe in, and not listen to the market or VCs too much. Think about who you're doing it with as seriously as you'd think about a relationship, because you're going to spend more time with your cofounders than with your own family. And if you're building an AI company, come to San Francisco and build an in-person team — this is where the renaissance is happening right now. Hire slow, hire great people; it pays dividends, because the people you bring in see who they'd be working with and want to be part of it, and it actually gets easier to close great people the longer you've been building the team, even if it means spending weeks wondering why nobody good enough is coming through the pipeline. Don't lower the bar.
Links & references
- Wordware ↗
- Wordware Sauna Waitlist ↗
- Wordware is hiring ↗
- Filip Kozera ↗
- Robert Chandler ↗
- Oana Olteanu ↗
This episode is brought to you by WorkOS.