Perspective

Nov 21, 2026 · Nugawi Intelligence · Perspective · 7 min read

The Software Factory Is Real: Our Production Numbers from 30 Days

The Software Factory Is Real: Our Production Numbers from 30 Days

The way software is built has changed more in the last two years than in the previous twenty. AI agents now write most of the code, the tests and the documentation, and the craft has moved to designing the system around them: the context they read, the guardrails they work within, the checks that verify their output, and the few decisions where a human really matters. The industry has settled on a name for that system: the software factory.

Most of what has been written about software factories so far is theory: concepts, predictions and product announcements. This post is not theory. Nugawi, an applied-AI consultancy in London, builds and operates a software factory in production for a public and social housing platform (we do not name clients) that serves 13 housing operators and programmes. Public housing is one of the most demanding domains software can touch: tenants, contracts, incidents and public money, every day. Here is what our factory produced in the last 30 days, measured from our own delivery systems.

The factory in 30 days

  • 1,545 changes merged into our codebases
  • 3 engineers directing the work
  • 28 of 30 days with an update reaching production
  • 2,300+ automated pipeline runs
  • 500+ backlog items delivered
  • 13 housing operators and programmes served from one platform

Web, API, Android and iOS apps, tenant portals, dashboards and AI features, all moving every day. Three people can only do that because the people are not the assembly line. They run it.

What a software factory actually is

A software factory is not a tool you buy, and it is not the agents. The agents are the assembly line. The factory is everything around them: captured intent, shared context, verification, shipping discipline and operations, arranged as a continuous loop.

Intent → Design → Build → Verify → Ship → Operate → new intent

Agents carry the volume through every stage. Humans set direction, approve the few decisions that carry real risk, and keep improving the factory itself. Every stage leaves a written, versioned trace that the next stage, human or agent, can read.

In a traditional team, engineers spend most of their day typing code and waiting for reviews, and a release is an event. In a factory, the loop runs daily and the release is a summary you read over coffee. The bottleneck moves from writing code to deciding what is worth building, and proving that what shipped actually works. That second half, deciding and proving, is where we think the engineering job now lives.

How our factory works

1. Capture intent automatically. Client meetings, emails and calls are turned into structured, prioritised tickets by AI. After one client working session, 63 findings were transcribed, classified and turned into 27 ready-to-work tickets the same day, each tagged by client, domain and priority. Nothing gets lost between the meeting room and the backlog.

2. One owner per domain, end to end. We do not split work into front end and back end queues. Each engineer owns a whole business domain, from the database to the mobile app, and directs a team of AI agents across every layer. No hand-offs between silos, no waiting for another team.

3. Shared context that every agent reads. Platform knowledge lives in versioned files, not in people’s heads: architecture, conventions, tenant rules, security policies and the lessons from past incidents. The overview of the whole platform is checked against the live systems and updated in the same change whenever something new is added. Every agent starts every task knowing how the platform really works.

4. Build in parallel. Agents plan before they code, implement across layers, write the tests and verify their own work. Engineers review the plan, steer and decide. Work that used to fill a sprint is measured in hours.

5. Ship with full traceability. Every change is linked to a ticket, and the ticket is updated automatically when the change reaches production. An evening release summary goes out by layer, mobile, web, API, database, and by feature, so everyone knows exactly what changed that day. That replaces approval queues that slow teams down without making them safer.

6. Operate and learn. Production errors become tickets automatically, with the context needed to fix them. A daily operations digest covers performance, capacity, security and costs, and every signal arrives with a what-to-do. Continuous threat detection watches the infrastructure. What happens in production feeds straight into the next round of work, and every incident becomes a rule that all agents follow from then on. The factory gets better every week.

Agents that work as a team

Our engineers’ agents hand work to each other. When a task belongs to someone else’s domain, an agent passes it, with full context, to that teammate’s agent, and the whole conversation stays attached to the ticket. Work moves across the team without meetings, without lost emails and without anyone re-explaining the problem.

The factory does not stop at our own team. Authorised client users can ask their own AI assistant questions about their data in plain language and publish dashboards from the conversation. One client’s team ships changes to their own internal system simply by talking to their assistant: each delivery passes automatic policy checks before it lands in pre-production, and the client promotes it to production when ready. Hundreds of automated delivery runs a month, with no engineer in the middle.

Where people stay in charge

A factory is not the absence of people. It is people placed where their judgement matters most. Humans decide:

  • what to build and in what order, together with the client;
  • anything that touches production data, cost or client commitments;
  • architecture and security, the decisions that are expensive to undo;
  • every conversation with a client.

Everything else, the volume, is done by agents and checked by the system. If that division of labour sounds like governance by default, it is.

Guardrails, because speed only counts if you can trust it

This factory runs for organisations that handle personal and public data, so trust is designed in, not bolted on:

  • EU data residency. The AI features run on EU-located infrastructure; data does not leave Europe to be processed.
  • Policy as code. Automatic checks stop deliveries that add unapproved software or new outbound connections. Servers only reach what they explicitly need.
  • A complete audit trail. From the request, to the ticket, to the change, to the deployment: everything recorded and traceable.
  • Aligned with Spanish and European rules. Controls designed around the Esquema Nacional de Seguridad (ENS), the GDPR and the EU AI Act, the same residency-first thinking behind our EU sovereign AI pipelines.

What this means if you are buying one

The platforms keep improving, but a software factory is not a product you install. It is your delivery system: your captured intent, your context, your guardrails, your verification, your decision points. Building that system is engineering-method work, which is what we do:

The engagement behind the numbers above is our housing case study, anonymised like all our client work.

What comes next

The factory keeps compounding. In build right now: independent verifier agents that check every change against our security, privacy and quality rules before it merges; scenario-based testing per client, so behaviour is proven rather than assumed; autonomous execution of routine work from ticket to tested change, with a person approving the result; and requests arriving from WhatsApp and email straight into the backlog.

If you want a factory without building one from scratch, we bring the method, the guardrails and a team that has already proven them in production. Start with a Discovery Sprint.

← Back to Blog