START HERE
What you will learn
- Explain a software factory without assuming full automation.
- Identify where repeatability helps a small team.
- Design a first delivery lane with evidence and human ownership.
Before you start: For stakeholders, independent builders and technology leaders. Bring one workflow or software change you can describe.
What is a software factory?
In this series, a software factory means an organised delivery system: specifications enter, bounded implementation work runs, independent checks produce evidence, and an authorised person decides what can be released. The term is used in different ways across the industry; this is an operating model, not a certification or a particular product.
An agent is one replaceable worker inside that system. It can investigate or implement a task, but it does not define business success, award itself permissions or decide that its own output is safe. A small team can use this approach with one repository and one worker. It does not require a fleet of agents.
Why build the process rather than collect more coding tools?
Generating changes faster can move the bottleneck into review. If nobody can explain the intended behaviour, recreate the environment or check a permission boundary, producing more code increases the amount of uncertainty. A repeatable process makes the handoffs inspectable: what was requested, what changed, what was tested and who accepted the result.
Start where work repeats: a document import, an integration adjustment or a well-understood bug class. Measure accepted changes and review effort, not generated lines or pull-request volume. The factory is valuable only if finished work improves without making failures harder to detect.
How the pieces connect
Human-led specifications describe the outcome and constraints. Version control records approved definitions. The runtime enforces the worker’s environment and authority. Tests examine both the delivered software and the agent’s behaviour. Release records tie the result to the versions and checks that produced it. Each following guide explains one part.
This is a conceptual delivery pattern, not a claim that every MintBios platform capability is generally available. Keep the production gate separate from passing a local test. A green run may justify review or staging; it does not automatically justify release.
Step-by-step process
Choose one delivery lane
Pick a bounded, repeated task with a named business owner and reviewer. Record the current lead time, review effort and failure modes.
Write entry and exit criteria
Require an approved specification, safe test data and scoped access before work begins. Define the evidence needed for review and the conditions that stop a run.
Connect the evidence
Attach the source revision, runtime image identity, test results and review decision to one execution record. Put definitions in Git and bulky evidence in protected storage.
Pilot before widening authority
Try representative tasks, including failures. Expand scope only after examining rework, blocked runs and recovery—not merely completion speed.
FOLLOW ONE EXAMPLE
Worked scenario
A small software team standardising import changes
A fictional four-person team regularly adjusts supplier CSV imports. Each engineer previously interpreted tickets differently.
- The owner defines how missing columns and duplicate records should behave.
- A worker changes one importer in an isolated checkout using synthetic files.
- CI checks valid imports, malformed inputs and unchanged existing behaviour; a person reviews the diff and evidence.
What this shows: The team gains a consistent acceptance path. It has not proved that every integration can be automated or that human review can be removed.
The process at a glance
Read Mermaid diagram source
flowchart TD
A[Human specification] --> B[Versioned delivery contract]
B --> C[Isolated agent runtime]
C --> D[Independent checks]
D --> E{Evidence acceptable}
E -->|No| F[Bounded repair or human review]
E -->|Yes| G[Human release decision]
F --> BWhat you should have at the end
- A delivery-lane definition with owned gates.
- A baseline for review effort, rework and accepted outcomes.
Common mistakes and how to avoid them
Calling any agent loop a factory
A loop without entry criteria, independent checks and release ownership repeats uncertainty. Add the decision boundaries first.
Treating throughput as the outcome
A large number of proposed changes can hide a growing review queue. Measure accepted, useful work and recovery effort.
APPLY THE IDEA
Try it yourself
Your task: Map one repeated software change from request to accepted release.
What to produce: A one-page lane map listing input, worker authority, independent checks, release owner and stop conditions.
Check your understanding: Can a colleague identify the person who may reject the result and the evidence they would use?
Key terms in plain English
- Delivery lane
- A defined path for one class of change.
- Acceptance evidence
- Recorded checks supporting an acceptance decision.
- Release gate
- A condition and authority required before deployment.