AI-Native Engineering 08: Build the Software Factory | Daniel Oduonye

This is part 9 of 9 in the AI-Native Engineering series. Start with the course index or continue to the next part below.

Audience. You should know one programming language, Git, a shell, HTTP, and how to run tests. The examples use plain files, JSON, HTTP, Git, and shell so the ideas transfer across agent SDKs.

Shared capstone. Every part extends the same shop-agent: a support agent that searches products, inspects orders, and prepares—but never executes—a refund recommendation. The companion repository lives alongside the site source.

Build the Software Factory

Capability. Turn a backlog into a safe, observable conveyor belt where background agents make incremental progress and humans steer.

A software factory is not “let the model write everything.” It is a control plane:

flowchart LR
    I[Issue with acceptance tests] -->|queues| Q[Queue / scheduler]
    Q -->|leases| W[Isolated worktree + agent]
    W -->|produces| V[Tests, lint, evals, security]
    V -->|gates| PR[Review-ready change]
    PR -->|requests| H[Human or agent review]
    H -->|fix| W
    H -->|merge| D[Deploy + telemetry]
    D --> G[Doc gardening + regression case]
    G --> I

Text equivalent: An issue with tests enters a queue. A worker receives an isolated worktree, produces code and evidence, and passes automated gates before review. Review either sends the work back for repair or merges it; deployment telemetry creates documentation and regression work that returns to the queue.

Module contract. Starting state: all prior artifacts in course/shop-agent. Add one issue with acceptance tests, risk, owner, checkpoint, and rollback; run both commands; prove the progress handoff is sufficient for a fresh agent.

Factory contracts. An issue has a measurable outcome, constraints, test cases, owner, and risk class. A run has a lease, budget, timeout, heartbeat, checkpoint, and cancellation. A merge requires tests, review evidence, and a clean diff. A deploy has rollback and post-deploy checks.

Measure the factory as a product: cost per successful outcome, time to first useful artifact, review acceptance rate, escaped defects, flaky-test rate, and human minutes per shipped change. PR count alone is a vanity metric.

Long-running design. Separate session (append-only record), harness (loop and routing), and sandbox (compute/files). Persist plans and progress outside the context window. Rehydrate from the last verified checkpoint, not from a model’s memory.

Entropy control. Schedule doc gardening, dead-code cleanup, dependency updates, flaky-test repair, and architecture checks. Throughput without cleanup eventually makes the repository less legible and slows every future agent.

Capstone acceptance. shop-agent can:

discover and load the right skill;
retrieve only relevant context;
call read-only tools through the gateway;
refuse injected instructions and wrong scopes;
survive a timeout without duplicating work;
produce a trace, tests, and review-ready diff;
resume after a context reset from its progress file;
explain what remains for a human.

Continue the series

← Previous: Parallelize Background Agents as a Bounded DAG

Return to the course index.