Wink Pings

Software factories sound like assembly lines—so why aren't developers buying in?

Zach Lloyd says software factory technology is hard, but the harder part is getting developers to want to use it. The replies push the problem further: trust, verification, secrets management, and risk concentration—none is easier than building the system.

Warp founder Zach Lloyd posted an article on X about putting software factories into practice. His judgment is direct: building a software factory is technically challenging, and people who claim it can be done easily probably aren't actually doing it. But the bigger challenge may be getting developers to want to use software factories—and to want to program as factory engineers.

First, correct a metaphor

Zach Lloyd thinks the problem starts with the term 'software factory.' It brings to mind mechanized, non-creative assembly lines and identical goods rolling off the line. Developers did not sign up for that. If the vision a team hears is going into a factory to tighten screws, they naturally won't be excited.

The alternative picture he offers is: in the future, engineers will be responsible for designing the factory, rather than standing beside the production line doing repetitive operations. This factory produces custom software, delivered to specific customers. The word factory only emphasizes throughput, cost, quality, and continuously improving those three things.

The really hard part is how people work

The second problem is more concrete. Developers have always had their own local cockpit: terminal, IDE or vim, shortcuts, and theme are all part of personal productivity. In the agent era, the equivalent is each person choosing their own toolchain, models, skill packs, and external system connections.

In the personal development tool paradigm, work happens privately first and only becomes public when a PR is opened. Software factories change this: working in public on the cloud by default. Zach Lloyd thinks this is good. Every developer managing coding agents in their own way does not scale.

His argument leans on management reality: VP Eng, CFO, and CISO may not care whether you use emacs or vim, but they will care which model you chose and how many tokens you spent; whether an insecure skill pack gives a prompt-injected agent your permissions; and whether someone on the team is 100x more productive.

The solution is to move the agent system from personal laptops to cloud infrastructure. Agent behavior is owned by the team, defined in code, continuously engineered, and improved collectively. The cost is taking away some developer autonomy and forcing them to work in public. Senior engineers will say local configuration makes individuals more productive. Zach Lloyd suggests acknowledging the friction, acknowledging some people will temporarily feel set back, and then painting the end state: shared agent infrastructure belongs to the team and organization; it should not be managed by a hundred people in a hundred different ways. Engineering leadership has to hold its ground when necessary.

Sharper questions in the replies

The replies did not stop at the technical level. Dusan says this is a hard sell; nobody signed up to tend the factory. Gabriele says nobody dreams of becoming a factory engineer. Sachin and Smobile both think getting the team to trust it is the harder half.

Martin mentions local setups can scale too. His bot builds trust through human approval gates and rollback backups, while shared infrastructure concentrates risk. Eslam offers a reverse perspective: a senior engineer's tuned local setup is also the factory's R&D lab; much of what goes into a shared config first came from someone's skill or MCP tweak.

webmcpx suggests using comparative experiments to drive adoption: run the same class of bug fix in a local setup and in the shared factory, track review minutes, retries, rollbacks, and token costs, and let engineers inspect failed runs and improve shared recipes. Gaurav says the verification layer is key; any failure is the factory's fault, which is an engineering problem with heavy responsibility. Vatsal says the harder part is knowing the factory's output is wrong before the customer tells you. Matt worries about secret sharing and authorization: a VM carrying all the secrets needed to drive a project can ship quickly, but it is high-risk and non-reproducible. Nour says their agent can draft on a Mac but cannot post or spend money; such hard limits actually make people trust it. Luke lays out the evaluation problem: handling ambiguous intent, the design phase, and getting input when needed; the harness is the capability around the model, and the factory is the methodology around the harness.

Iris Millie posted several images in the replies and thinks this angle holds up in the bigger picture:

image

image

image

image

In this discussion, the hard part of software factories shifts from building the system to organizational trust and verification. Zach Lloyd's end-state judgment is clear: the future is shared agent infrastructure, with coding in the cloud, guardrails, and collective measurement and improvement. The doubts in the replies are also clear: sharing concentrates risk, personal setups have R&D value, and trust requires approvals, rollbacks, and verifiable results. Whether adoption happens depends on whether teams can treat these problems as engineering problems and solve them together.

发布时间: 2026-11-17 03:31