Contents

THE AI-NATIVE SOFTWARE DEVELOPMENT PLAYBOOK

Chapter 02

Five industry approaches, one accountable workflow

Where Anthropic, AWS, Atlassian, Google and OpenAI converge—and where they do not.

Compare the sources before choosing a method. Synthesis is a decision, not a list of quotations.

Five lenses on the same change

The five publishers agree that AI changes more than typing speed. Their emphases, however, are different. Anthropic describes modular, stage-linked practices with versioned artifacts, governance gates and an eventual goal of accepted artifacts triggering subsequent steps; AWS proposes the AI-DLC methodology with Inception, Construction and Operations; Atlassian foregrounds connected organisational context and measurement; Google explores the shift from vibe coding to disciplined agentic engineering; and OpenAI describes delegation, review and continuing engineering ownership. Automated handoffs are an intended end-state, not a prerequisite for adopting individual plays. The table below is our editorial synthesis, not a vendor-issued scorecard [1–6].

OrganisationDistinctive emphasisEditorial application
AnthropicVersioned artifacts and lifecycle gatesAutomate suitable gates with evidence; reserve consequential decisions for accountable people [1]
AWSAI-DLC inception, construction, operations; boltsKeep the useful decision cycle; adopt jargon selectively [2–3]
AtlassianContext, measurement, operating modelFix shared context and delivery metrics before scaling [4]
GoogleVibe coding versus disciplined agentic engineeringKeep architectural judgement and validation visible [5]
OpenAIDelegate, review, ownAgents draft; engineers retain ambiguous and critical ownership [6]

Three tensions worth resolving

  1. New rituals versus established delivery methods. AWS gives teams a distinct AI-DLC vocabulary, including “bolts”, short, intensive delivery cycles measured in hours or days rather than weeks [2]. Atlassian emphasises transforming the operating model, while Anthropic offers modular plays that need not be adopted simultaneously. Our judgement: use the artifacts and approval disciplines that improve outcomes; do not rename sprints or replace established Agile/DevOps ceremonies unless the new structure solves a demonstrated coordination problem [1–4].

  2. Automated decisions versus accountable decision rights. Anthropic describes increasingly autonomous development workflows, while retaining human accountability for judgements that matter. The meaningful design question is which gates a machine may close, which require a named person, and what evidence each decision must leave. Accepted artifacts may eventually advance stages automatically, but automation should neither silently rewrite policy nor assume authority that was never delegated. In ServiceFlow, an agent raised the resource-existence disclosure question; Justin Underhill made the 404-versus-403 security decision before implementation [1,6,10].

  3. Agent-level guardrails versus system-enforced guarantees. Anthropic distinguishes advisory skills from deterministic hooks and recommends checks at repository boundaries; Atlassian and Google likewise stress reliable context. Hooks can constrain agent actions in a configured environment, but application permissions must hold regardless of the agent or person submitting a request, and repository merge rules must hold regardless of who pushes a change. ServiceFlow enforces request-level authorisation in server-side API code, tests the refusal paths, and requires passing CI through an active GitHub branch ruleset. Tests verify the control; they do not replace it [1–5,10].

Our synthesis: intent → context → bounded execution → independent evidence → accountable decision → operational feedback. This six-part loop organises the practices discussed in this playbook into a practical sequence for teams to tailor and test against their own delivery needs.

How to apply this chapter

Choose one bottleneck and a measurable hypothesis before selecting a tool. If requirements ambiguity is dominant, start with intent and specification review. If changes ship but break, prioritise negative tests and enforced release checks. If agents cannot find reliable knowledge, invest in repository context. Only expand autonomy after the team can demonstrate evidence of correctness and reliable rollback.

Research basis: Anthropic [1], AWS [2–3], Atlassian [4], Google [5], OpenAI [6].

top