AI strategy: Software developer

Prioritized AI adoption map. Date: 2026-06. Sources: GDPval, AEI, O*NET, and profession-automation-2026.md. ← Role hub

Context
Computer and mathematical work is among the task classes with the greatest AI exposure in the Anthropic Economic Index, which is only a guide. Coding assistants are already common. Junior tasks are moving first. Architecture, security, and production responsibility are becoming more valuable.

1. Code production (high priority)

No.	Goals	KPIs	Initiatives
1	Boilerplate, tests, and docs	time · test coverage	Copilot or Cursor · human review
2	Agentic tasks across multiple files	feature time · share of accepted diffs	Cursor Composer/Copilot agent · acceptance criteria

2. Quality and review (high priority)

No.	Goals	KPIs	Initiatives
3	PR summaries and automated review	review time · defects found	AI comments · final human review
4	Security scans and migrations	vulnerabilities · migration time	Amazon Q or scanners · automated transformations

3. Understanding systems (medium priority)

No.	Goals	KPIs	Initiatives
5	Explain and navigate the codebase	onboarding time · debugging speed	prompt: "Explain this module and its risky areas"

4. Data security (foundation)

No.	Goals	KPIs	Initiatives
6	Keep private code from leaking	incidents	local models such as Ollama · never paste secrets

5. The role (foundation)

No.	Goals	KPIs	Initiatives
7	Become an "agent boss" and own architecture	% of time on architecture and review	orchestrate agents · own production responsibility

Where to start: the first round

The tables are a map of what is possible, not your plan. Start with process scoring. Ask whether the task repeats, has enough volume, produces a checkable result, carries an affordable error cost, and has usable data. For most developers, the first round looks like this:

No. 1, boilerplate, tests, and documentation. It repeats on every task and has enough volume. Run the tests instead of taking the output on faith.
No. 3, PR summaries and draft review comments. An error is cheap because a person still performs the final review. AI only speeds up the first pass.
No. 5, code explanation and navigation. The error cost is low because this is guidance for a person during onboarding or debugging, not code going to production.

Not in the first round: No. 2, "agentic tasks across multiple files," has a higher error cost because changes spread across several files. Build review discipline on Nos. 1 and 3 first. No. 4, "security scans and migrations," can miss an expensive vulnerability. Trust it only after the review practice in No. 3 is reliable. Never send secrets or private code from No. 6 to third-party services. This is a boundary, not a scheduling choice. See the red flags in the role hub.

Keep these parts of software development human

Architecture and system boundaries. AI can offer options; a person decides.
Final review before production. It cannot be delegated at any maturity level.
Secrets, keys, and private code. Do not put them into third-party services. Use local models for private material.
Responsibility for a production incident. A person owns it even when an agent wrote the code.
Licenses for generated code. A person checks provenance and compatibility.

Review points and stop thresholds

Give every first-round initiative its own review point in the Human Review Matrix and set a threshold in advance in the agent contract:

Initiative	What a person checks	Stop threshold (example, replace with your own)
No. 1 boilerplate and tests	run tests and inspect edge cases	if a test fails or an edge case is missing, finish it by hand and do not merge
No. 3 PR review	a person completes final review of the diff	a missed bug or vulnerability means revising the review-bot prompt and rechecking recent PRs
No. 5 code explanation	compare the explanation with real behavior in at least one example	if the explanation differs from the code, do not use it for onboarding without verification

Before you trust the flow, build a reference set of 20 cases: 20 reviewed PRs or tasks with known correct diffs. Test a new agent or prompt on them before using it in the live task flow.

Two paths from here

You are a developer: open the role hub, map your week, and build a personal plan in the Volume 1 workbook.
You lead engineering: use process scoring, write the first agent contract, and complete the full path in the Volume 2 workbook.

Discipline
Send code to production only after review. Never paste secrets. A person owns the architecture.

→ Role hub · audit · plan.

Sources: profession-automation-2026.md; book, Chapters 2 and 3. Version: 2026-07-09. What changed: added "Where to start: the first round," "Keep these parts human," "Review points and stop thresholds," and "Two paths from here."