Contents

THE AI-NATIVE SOFTWARE DEVELOPMENT PLAYBOOK

Chapter 05

The AI-native development workflow

Bound the work. Verify the evidence.

Follow a change from approved requirement to controlled production release.

A controlled loop, not an uncontrolled handoff

A repeatable workflow should turn approved intent into a small unit of work, provide approved context, let an agent attempt the task, verify the result independently and preserve evidence for the next decision. The workflow continues after release through monitoring and improvement.

StepAgent may doResponsible human or control
FrameSummarise requirements, propose gapsProduct owner approves intent and scope
DesignSuggest options and trade-offsEngineer/architect validates architecture
ImplementChange scoped files and draft testsRepository and permission boundaries enforce scope
VerifyRun unit, integration and static checksCI checks run independently; reviewers inspect risk
ReleasePrepare release notes and deployment planAuthorised workflow approves production change
OperateSummarise incidents and propose remediationsOperations owner controls interventions

An exemplary work packet

A ServiceFlow feature task could state: “Implement manager approval of requests, using the existing role mapping. Do not alter database permissions without review. Add tests for self-approval, cross-department approval, invalid requests and audit history. Open a pull request with changed files, test output, unresolved assumptions and a rollback note.”

The agent can inspect relevant code, implement a bounded change, execute tests in an isolated environment, and prepare a pull request. It must not be allowed to infer production credentials, turn off a failing security check, or self-approve its release.

Evidence is the unit of progress

Every change should have traceable intent, code diff, tests, checks, review disposition and release record. This chain is not automatically trustworthy because it exists: test quality, coverage, reviewer independence and the integrity of tool outputs must also be examined.

Parallelism needs limits

Multiple agents can tackle separable tasks, but shared file ownership, competing changes and dependency conflicts can erase the benefits. Give parallel work explicit file boundaries, interfaces and an integration owner. Begin sequentially for high-risk changes or tightly coupled components.

IN PRACTICE Require every delegated task to return a change summary, evidence, remaining uncertainty and proposed next action; never accept “done” as the only result.

Research basis: Anthropic [1], AWS [2–3], Google [5], OpenAI [6].

Branch and evidence discipline

Prefer an isolated branch or workspace. Require a diff, concise explanation of changes, tests executed, tests not executed, known limitations and any dependency additions. Failed tests are findings to investigate, not invitations to remove the test. Repository rules should prevent bypassing required checks, even when a coding agent recommends doing so. For larger tasks, checkpoints and human reviews are cheaper than discovering architectural divergence at the end.

What a genuinely completed change looks like

For the ServiceFlow approval endpoint, completion means: request and approver identity are resolved server-side; unauthorised approvals are rejected; repeated submissions are handled safely; audit events are recorded without leaking sensitive data; a test suite includes negative cases; the reviewer confirms the business rule; and the release process includes monitoring and recovery. A visually convincing demo is not the acceptance criterion.

top