A software factory that shows its work

AI can produce a lot of code. That isn't the same as delivering a change you can trust.

StratoFusion's software factory workflow is about closing that gap. Give an agent a clear task, room to work, checks to pass and evidence to bring back. The developer decides what ships.

We call it the Agent Delivery Harness. The name matters less than the result: a handoff you can inspect instead of a confident “done” you have to investigate.

From a brief to a reviewable change

The workflow puts five habits around AI-assisted development:

1. Agree the job. Define what “done” means and identify sensitive boundaries before implementation starts.
2. Keep work separate. Use an isolated Git worktree for each task. Declare shared ports and test resources so conflicts can be spotted before work starts.
3. Bring back proof. Attach the diff, relevant test results and useful visuals. Show what changed and what was actually checked.
4. Review, then repair. Use fresh-context review for sensitive work where practical. Limit repair cycles to two, then stop and ask for help.
5. Let a person ship it. Keep CI and security checks in place. Merge to development, then make a separate, human-approved release.

For a solo developer, the appeal is practical: less detective work at review time, clearer stopping points and smaller changes to accept, revise or roll back. Agents do the implementation work without acquiring unattended authority over production data, credentials or destructive operations.

Small teams can use the same workflow: a shared definition of done, clear task ownership and consistent review handoffs. Shared resources still need coordination; the pilot didn't measure team throughput.

What the five trials taught us

We tried the approach on the workflow tooling itself, streaming documentation, provider-architecture documentation, a shared toolbar refactor and a survey-results refactor.

Different jobs needed different evidence. Documentation benefited from source inventories and link checks. The refactors needed tests showing that existing behaviour survived the cleanup—not screenshots of an unchanged screen.

Four trials received usefulness feedback: three “yes” and one “partly”; one response was missing. All came from me as the sole maintainer, so this wasn't an independent productivity benchmark. No confirmed privacy disclosure occurred, and CI, security checks and human approval stayed intact.

The decision was adopt with revisions: keep the handoffs concise, strengthen evidence safety and bound the repair loop. The five-trial scorecard holds the detail.

Make the handoff visible: the public-page redesign

The later redesign gives us a concrete example of the evidence this workflow values. It covered the homepage, welcome, about, article and policy pages: shared navigation, responsive layouts, light and dark themes, and nine animated product demonstrations.

The redesign did not opt into the harness; it is an example to learn from, not a sixth pilot trial.

For review, the question becomes specific: does the new presentation explain StratoFusion more clearly? Put the result in front of the reviewer.

Before: the original homepage message.

After: a shared visual style with the product at its centre.

Motion adds another layer of evidence. These search and copy demos make the delivered presentation easy to inspect: a query becomes results; a source and destination become a visible sequence.

The PR recorded 94 focused tests, layout checks at four widths in both themes, and 15 browser interaction checks. The visuals show the presentation; the checks test its behaviour. Demo timing isn't a measurement of live transfer speed.

That is the pattern to carry into the next task: a clear request, a visible result and checks matched to the change.

Why no orchestration platform?

We needed a repeatable workflow before we needed an agent fleet.

Git, pull requests, tests and security checks were already in place. Repository rules and small helpers could connect them without another service to host, configure or maintain. With an initial limit of two implementation worktrees, explicit ownership and human coordination fit the job.

An orchestration platform may help later with scheduling many tasks and managing dependencies. It doesn't replace a clear brief, useful evidence or a release decision.

The inspiration came from a software factory video and its companion skills. The video was Cursor-sponsored; we adopted useful ideas, not its speed claims.

Better handoffs. A developer still in charge.

The goal isn't more autonomous activity. It's useful work that arrives ready to review.

That is StratoFusion's software factory: agents help build, evidence makes the work inspectable, and people keep the final say.

Explore the delivery workflow or see the public-page example.