The Agentic Software Factory

A software factory, in its older sense, is a repeatable delivery system: standardized inputs, a defined build path, automated quality checks, and a predictable output every time. The agentic version keeps that shape and replaces the workers inside it: fleets of AI coding agents write the code, run the tests and iterate, while engineers define the intent for a change and review the evidence the agents produce. BCG Platinion describes this model as a system that takes a signal in at one end, a bug report or a business requirement, and produces deployed, verified software at the other.

A bank considering this model is not buying a single product. It is restructuring how a development organization spends its engineering time: less on writing routine code by hand, more on the harness engineering that keeps agent output reliable, and on reviewing what the agents produce before it reaches production.

The clearest early use case inside banking is a migration: moving a service off a legacy platform or a mainframe, where agents handle the mechanical translation work under a developer's review, a task covered in more detail on the legacy and core banking modernization page. The same pipeline applies to writing test coverage for an existing system or keeping documentation in sync with code, work that previously competed for a developer's time against new features.

The review gate stays in place. A bank still needs the same proof of who approved a production change and why, whether a human or an agent wrote the code. The agentic software factory moves work from typing to reviewing; it does not remove the review.

A software factory is a repeatable system for turning a requirement into deployed software: standardized inputs, a defined assembly path, and automated checks on the output. The agentic form of the idea puts AI coding agents inside that pipeline as the workers who write and test the code, with engineers setting intent and reviewing results instead of writing every line themselves.

Harness engineering is the discipline of building and tuning the scaffolding an AI coding agent runs inside: its tools, its permissions, and the checks on its output. An agentic software factory runs on many agents at once, so a weak harness multiplies its own mistakes across every agent using it. The full explanation, including why the discipline is still hard to plan for, sits on the AI coding agents page.

No. It changes what a developer spends time on: less time writing routine code, more time defining what a change should do and reviewing whether the agents got it right. A bank's platform team still owns the audit trail, the test gate and the deployment rules; the factory changes who produces the first draft of the code those rules check.