Justin McCarthy—who led the StrongDM Software Factory work before founding Diffusion—opened his summit talk with a deliberate provocation aimed straight at teams rushing to automate unproven ideas.
“So that's of course the time for me to say, please stop building software factories.” Justin McCarthy · Brownfield Factories · 00:02:12
When product-market fit is still an open question, human spec writers routinely ask for "a faster horse" because nobody yet knows what the software should do. Autonomous factories simply burn tokens hardening the wrong assumptions. Brownfield systems invert that geometry: an established codebase already has customer telemetry, support tickets, production traces, and revenue proving what works.
“It's something where customers are saying with their dollars and their and their utilization. Like yeah, this works for me. I like it. I want more of this.” Justin McCarthy · Brownfield Factories · 00:03:18
From the commercial product side at Lovable, Fabian Hedin reached the same boundary condition from a different direction: code is never the asset a factory optimizes for on its own—live business loops around payments, analytics, SEO, and customer outcomes are.
“But in the end, what matters is actually not the code.” Fabian Hedin · What it takes to set intelligence free · 00:02:04
Once brownfield specifications are legible to coding agents, coordination overhead—not typing speed—becomes the main drag on throughput. Quinn Slack described spinning out a 20–21 person team under an external hiring freeze, shedding departmental silos, mandatory PR queues, PM layers, and 96% of external SaaS seats in favor of internal mini-apps running in agent sandboxes.
“And that skin in the game that you get when AI makes it so you can operate with way less people. Skin in the game—it's like this spiritual thing for the soul. Of great people, it makes them better at everything they do.” Quinn Slack · A Company Built Around the Agent · 00:05:51
Slack grounded the cultural claim in hard financials: eliminating legacy enterprise sales-and-process checklists let a small team compete directly at massive capital efficiency.
“That is profitable. That's more than a hundred million ARR. That is doing so. On a dollar in, dollar out, you know, basis, that's 40 to 50 times more efficient than old-fashioned companies.” Quinn Slack · A Company Built Around the Agent · 00:15:58
Walden Yan reported a parallel restructuring at Cognition, where merged pull requests grew 40× on 2.7× headcount after shrinking teams into two-to-three-engineer pods that own everything from implementation to triage and customer issues—while Raz Berman Gold (Overcut) cautioned that enterprise factories fail ROI whenever they require months of Forward Deployed Engineers rather than plugging cleanly into existing repos and trackers.
“But they're responsible for everything beyond engineering as well.” Walden Yan · A New Age of Software Engineering · 00:01:48
Behind Amp's "no token limits" posture sits a sharp economic warning from Steve Yegge. Running fleets of parallel agents quickly exposes who is actually paying the compute bill—Yegge estimated his 22 consumer Max plans ($10k/month) represented roughly $1.5M/year at raw API pricing—and panicked enterprises routinely break factories by clamping down too hard.
“but I tell you, every problem is a few problem.” Steve Yegge · Software Factories: Escape from Trash Mountain · 00:09:22
“That they put all these rules in place to make sure nothing bad ever happened, and so nothing ever happened. Okay, let's call it overfencing.” Steve Yegge · Software Factories: Escape from Trash Mountain · 00:15:27
Rather than throttling developers blindly, infrastructure and harness builders attacked waste directly. Natalie Serrino (Gimlet Labs) showed why the 63-second median coding-agent turn shatters focus and forces asynchronous drift. Philipp Schmid rebuilt his agent harness every six months to delete 90% of obsolete scaffolding, while Jarred Sumner kept Claude Code in TypeScript and cut memory ~4× with ahead-of-time JavaScript bundling.
At the tool and classifier layer, Arsalan Bashir (Sentry) replaced naive 1-tool-per-API wrappers with dynamic discovery and typed execution, running 4× faster with 30% fewer tokens. Scott Roberts (Workday) and Devin Gupta (Parallel) demonstrated cheap, calibrated decision models running inside SQL queries and CI triage for pennies per thousand judgments—cutting repeat CI-triage token consumption by 94% while reserving frontier tokens for genuine synthesis.
Multi-hour factory work breaks if every deploy, crash, or dependency install forces a cold restart. Lovable graduated from a hard 30-minute limit to 10+ hour runs across roughly a million sandboxes a day by isolating per-agent message inboxes and booting containers from copy-on-write golden checkpoints in a fraction of a second.
“all of them you copy it over in about 230 milliseconds.” Fabian Hedin · What it takes to set intelligence free · 00:11:24
Across the runtime stack, Michelle Chen explained how Cloudflare paired Workers, Durable Objects, and edge containers to close the remaining gap in its agent cloud, while Jarred Sumner noted that long-lived subagents outperform rigid step-by-step workflows precisely because they preserve working mental context across files. When given bare infrastructure without brittle orchestration, frontier models even construct their own persistence layers.
“So agents now are sufficiently advanced that they can actually build primitives if they need to solve hard problems.” Lance Martin · Exploring Claude’s frontier · 00:03:46
McCarthy's operational blueprint for brownfield codebases is deliberately unglamorous: wire agents directly to production reality so they continuously protect existing user contracts without waiting on sprint ceremonies.
“So you've got an existing codestate. You've got an existing set of promises. Just make it continuous. Make it loop and make it keep those promises.” Justin McCarthy · Brownfield Factories · 00:06:30
Yet keeping promises steady is table stakes once competitors run the same brownfield loops. McCarthy argued that durable moat shifts outward from delivery to demand creation—expanding customers, wallet share, SKUs, and product lines—echoing Mihail Eric's lesson from building Monaco's agentic go-to-market platform.
“We have to figure out how to build a demand factory. Because our brownfield factory can't build the moat. Right. All it can do is keep our existing promises running well.” Justin McCarthy · Brownfield Factories · 00:13:17
“Go to market is one of the few remaining sources of alpha” Mihail Eric · Building World-class Agentic GTM: Learnings from Monaco for Monaco · 00:05:43
While McCarthy pushed for 24/7 continuous brownfield loops, Walden Yan offered the most extreme long-horizon proof point: autonomous agents sustaining cross-platform parity over more than six weeks with minimal human steering.
“Go manage the entire Android application, and over the course of 44 days, we have a working application, which we'll be putting out soon.” Walden Yan · A New Age of Software Engineering · 00:10:45
Simon Willison paired that autonomy with a second counterpoint to elaborate factory architectures: once frontier models can solve hard problems almost by brute force over 12-hour test runs, defining the problem and verifying taste often beats elaborate multi-stage scaffolding.
“As difficult as possible, the models will solve your problem just almost with brute force.” Simon Willison · 2026 in LLMs · 00:08:02
Looking back after leaving the Claude Code team, Thariq Shihipar challenged the entire fire-and-forget asynchronous workflow. Delegating background runs breaks deep flow and leaves engineers ignorant of the system they are supposed to maintain:
“And I find that like the ball just makes a bunch of choices for me, and then I try it out. I'm like, wow, I have no idea what that actually does. I get distracted pretty easily. I don't feel in the zone very much. You know, I sort of send off a prompt and they start doing something else and I come back. And I honestly make less progress because of these things, right?” Thariq Shihipar · Thariq Shihipar, Independent · 00:11:06