In a recent episode of Pale Blue Nexus, Diffco CEO and Co-Founder Vadim Peskov sat down with host Yohann Calpu to talk about what agentic development has actually changed on real projects — and what it hasn’t. His side of the conversation, in his own words.
The 10x engineer number is a tagline, not a metric
Every conversation about AI and engineering lands on multipliers. 10x engineer. 100x engineer. Great headline, close to meaningless on an actual project.
Start with the obvious problem: how would you measure it? Same trap as counting lines of code. Output per ticket? Also questionable — in much of our work now, a human never touches the ticket or the code at all. You could measure development speed, but on most real projects the limiting factor was never typing. It was architecture and product decisions. Those are still made by people, and they still take the time they take.
So when someone says their team is 100x now, the useful follow-up is: 100x at what?
What actually changed: cost
Here’s a number we can see on invoices. We’ve cut project costs for clients by at least 4x.
That holds for most engagements, with one large caveat nobody puts in the headline: it depends enormously on whether the project is greenfield.
There’s a myth that you can point AI at a 20-year-old legacy codebase and watch wonders happen. That is not what happens. You inherit exactly the problems you always inherited — undocumented decisions, dead paths, coupling nobody remembers introducing — and you find them one surprise at a time. AI doesn’t dissolve accumulated entropy. It refactors faster once you’ve found it.
The ceiling is real too. Nobody is building a Salesforce or HubSpot competitor in a day, this year or next. Can you build a large system faster than before? Absolutely. Those are two different claims, and they get conflated constantly.
We stopped writing code
We’ve taken on multiple projects in stacks where we had limited in-house depth. We never wrote a line of code on them. We found real issues in legacy systems, fixed them, shipped them — all agentic. Not to make a point, but because writing it by hand was worse on every axis: slower, more expensive, and more error-prone.
That last part surprises people. Will AI make mistakes? Of course. But compare error rates honestly — AI against a human working the same deadline — and AI comes out fine.
On smaller new projects, one-month builds, we’re essentially at fully agentic development. Most of the calendar goes to discussing functionality and making product decisions. The building takes days.
Which leads to the rule that matters most: when something breaks, don’t touch the code. Fix the prompting, the harnessing, the memory, the documentation. You’ll have bugs and security issues — expected at this speed. Fix the system that produced them. The moment you hand-patch the output, you’ve broken the premise of agentic development and you’re maintaining code by hand again, with extra steps.
Architects, not coders
We don’t need coders, in the sense of people who write code, because we’re largely not writing code. We need people who can design a system — architects and product people — and then build the right harness for AI to construct it. Which language you know matters far less than how fast you learn.
Better models don’t equal better output
Model advancement is not moving output quality nearly as much as people assume. The practical difference between frontier models is minimal — if you have the right harness and if your documentation isn’t careless. Get both right and you can run open-source models at roughly a tenth of the inference cost for output that isn’t meaningfully worse.
The bottleneck isn’t the model. It’s the information going into it.
And “prompting” has become a lazy diagnosis. The problem isn’t phrasing. It’s that people don’t give AI enough information about what they need. Two lines — build me a website, I sell balloons for holiday parties — is not a description of a task. Before you ask for the site you need to have thought through the content and the user flows. Better models will not fix that.
Two predictions for the next 12–18 months
Open source wins on private data. More multi-agent systems will run on open-source models, specifically because you’re not handing your data to anyone. If you’re building another SaaS product, it may not matter. If you’re sitting on genuinely sensitive data, the right harness closes most of the quality gap and the inference economics aren’t close.
The interface stops being something you fill in. Not chat replacing UI — you still want visual representation. But no more clicking through a CRM to log what already happened. Instead: you spoke with John, I’ve logged it, here’s the proposal I generated, please review and approve; here’s the contract; here are three suggestions for the product team, already sent. I’d pay more for that CRM than the one I have. And you can build it inside HubSpot or Salesforce today, if you set it up right. It just isn’t the default.
🎧 Watch the full conversation:
