Example · SDET / Test Automation Engineer

63% of your week as an SDET is still yours.

AI now writes test code and fixtures. Knowing whether a failure is real stays with you.

Make my planA typical week at the middle pace. Your plan uses your own tasks.

Is test automation a dying career?

In a typical week at the middle pace, 28% of this work needs your judgment and 35% is work you do using AI. AI could do about 37%. New tests from a spec, page objects and fixtures, the pipeline config and the load-test script make up 39% of your week. Models already do this work well, and it is the part you were hired for.

Read more

If you have let an agent loose on your test code, you have seen it. It writes clean page objects in minutes.

The same experiment shows what holds. Some generated tests assert a bug into place because the spec was wrong. Some "fix" a flaky test by adding a retry. Catching both takes a person who knows how the system fails. Your job now is deciding whether red is real, deciding what a release gate should block, and reading a performance result against a system you understand. You already do all three, and your title says "engineer in test" as if the test were the point. For many SDETs, two of the three are the hardest parts of the week: the early-morning red pipeline and the request to make it faster. This report helps you put your name on the decisions you already make.

The week, task by task

Who does each hour of a typical 40-hour week.

See every task
You28% · 11 hours

Needs your judgment.

  • Performance risk judgment: load profiles, SLO thresholds, "is that p95 real"7%
  • Release gate ownership: what blocks merge, what is waived, the early-morning call7%
  • Test strategy and design reviews: "how will we test this?"6%
  • Production incident support: "was this testable, and why did we miss it?"4%
  • + 1 smaller task3%
You using AI35% · 14 hours

AI does the slow parts. You still own it.

  • Flaky-test triage and root cause: bug, test or environment12%
  • Reviewing tests: developers' and agent-generated PRs8%
  • Test environments and data: ephemeral envs, seeding, secrets, containers8%
  • Contract tests: consumer and provider pacts, broker, can-I-deploy6%
AI37% · 15 hours

AI could do most of it.

  • Writing new API and end-to-end tests from a ticket or spec12%
  • Framework code: page objects, fixtures, helpers, client wrappers10%
  • CI pipeline config: test stages, caching, sharding, matrix, YAML7%
  • Performance scripts: load scenarios, thresholds, data generators (e.g. k6)6%
  • + 1 smaller task4%
Why some tasks still need a person
"Is the nightly red because of a bug, the test, or the environment?"

The answer lives in the history of what broke, which services share a database, and which squad shipped late. A model can read the log, but it does not know that one service always times out on the first run after a schema migration.

Reading a load-test result and saying "that p95 is real" about the slowest 5% of requests

A performance drop looks the same whether the code regressed or the perf environment was under-provisioned that night. Calling it wrong either way costs a release or an outage. The judgment needs the environment's history and the business's tolerance, and neither is written down.

Deciding what the merge gate blocks

Where your company is audited, the merge-blocking pipeline is evidence in a change-management control. Someone has to be answerable for what it checks and what it waives. Often, informally, that is the SDET.

Knowing a generated test asserts the wrong behavior

A green test that locks in a bug is the most expensive kind of mistake, because it looks like coverage. The person who reads a test against what the system should do, rather than what the spec said, is the constraint on every test-generation tool.

Telling a service team their contract change will break a consumer before they deploy

Contract tests catch the break. Knowing which consumer matters, who owns it and whether the break is acceptable this sprint is organizational knowledge.

Good to know

What is hard about this field, and what you can do about it.

The hard truth

AI speeds delivery, but stability is lagging behind.

Teams using more AI tend to ship faster, and their releases tend to be less stable. Closing that gap is your skill set.

What you can doTrack change failure rate and rollback time on your team. Bring those numbers to planning, along with the tests and gates that would bring them down.
Sources (2)

What's next for an SDET?

Plan A is the best move inside the work you do now.

APlan A

Grow in the job you have

The main focus. Most weekly sessions build toward it.

Your next stepBecome the named release-gate owner
Each week
4 sessions
Ready in
About 5 weeks to the first gate standard
Why it fits

Why you. If you already make the call when the nightly is red, a record of telling real failures from noise is standing nobody else on your team has. The gap between what you do and what your title says is smaller here than for your other moves.

The gap. Usually no written gate standard exists, and waivers happen in chat. Nobody has named an owner at your level, so it can be anybody, including "pipeline green."

The trade-off. You become the person blamed, in writing, when a waived threshold becomes an outage. The stress stays the same, but you have more authority on paper. You will also ship fewer tests and less framework code for a quarter, and someone will count.

2 other moves that fit
  • Own performance risk decisionsReady in about 6 months to be the person who signs off on performance for every release
  • The portable hedge: test-infrastructure engineer with a public ledgerReady in about 10 weeks to know what the market pays

What's a plan B for test automation engineers?

Plan B is meant to be different from this job. It was picked because skills from this role carry over.

BPlan B

Get ready for a new kind of job

A second option. A newer job AI is creating that these skills already fit.

A job to aim forAI evaluation specialist

Your nightly suites already tell real failures from flaky ones, and building eval harnesses for AI features uses that same judgment and tooling.

Each week
1 session
Outlook
Hiring now
Pay
$230K to $270K a year (one posting)One Anthropic safety evaluations posting, 2026; AI lab pay runs well above most employers.
2 other options
  • Agent reliability and quality engineerOwning whether the build is actually red is reliability work, and a shipped agent test harness plus a few more years opens this senior door.
  • Agent simulation scenario designerScripting scenarios that drive a system through every path is what agent simulation needs, with conversations in place of page objects.

In your own plan you can switch to any of the 47 roles.

The 13 weeks

Small steps that add up. One plan for both: most sessions build Plan A, one a week builds Plan B.

By week 13 you will have

AFor Plan A

  • A signed release gate standard with a waiver log
  • A quarantine policy for flaky tests
  • A one-page memo asking to own the gate

BAnd for Plan B

  • A draft write-up of your AI test
  • A published write-up of your AI test for your portfolio
Month 1Weeks 1 to 4
  1. 1Measure your week
  2. 2Check the AI tool policy
  3. 3Trace escaped bugs
  4. 4Write a review checklist

You finish withA disagreement count and a revised checklist

Month 2Weeks 5 to 8
  1. 5Write a release gate standard
  2. 6Try mutation testing
  3. 7Set a flakiness budget
  4. 8Write agent rules and audit thresholds

You finish withAn audit of every load-test threshold

Month 3Weeks 9 to 13
  1. 9Update the standard and ask
  2. 10Check the job market
  3. 11Gather your evidence
  4. 12Make your ask
  5. 13Choose your path

You finish withOne application sent, if the answer was no

Three courses from the plan
  • ISTQB Certified Tester Foundation Level (CTFL) v4.0ISTQB via AT*SQA (US exam) · $229

    Only if the postings you want ask for it. It is the prerequisite for every other ISTQB paper.

  • ISTQB Advanced Level Test Automation Engineering (CTAL-TAE) v2.0ISTQB via AT*SQA · $249

    No. It certifies the exposed half, and a senior SDET usually already owns the architecture.

  • ISTQB Test Automation Strategy (CT-TAS) v1.0ISTQB via AT*SQA · $249

    Maybe after Week 12 if Plan A lands and you want the vocabulary for a strategy document.

Checked 2026-10-11.

The fuller picture

Risks beyond AI: highCost-center risk

This risk does not come from AI. Test automation is an engineering cost center and gets cut when engineering wants to look lean, because a green pipeline is invisible and a red one looks like your fault. Yahoo pushed to eliminate its separate QA function between mid-2014 and early 2015, before any of today's tools existed, and the 2022 to 2023 tech layoffs hit QA and platform teams again. Under the moderate scenario this risk is about the same size as the AI one for you, and the remedy for each is different.

Where this could be wrong
  • Flaky triage may be more exposed than we scored. Our capability score assumes agents classify well and root-cause badly. If a triage agent fed your thirty-run history correctly labels the environment failures in Week 1, that task becomes more exposed and owning the release gate becomes urgent.
  • We assumed the gate call is yours informally. If a senior SDET or a release manager already decides and you report, the release-gate move means joining the decision rather than owning it, and that task gives you less protection.
  • Your company may not care about performance. The performance move assumes a product where latency costs money. If performance results have gone unread for a year, that move is a hobby and the plan is the release gate and the hedge.
  • A senior SDET could claim the release gate fast. More experience and the framework in hand make a stronger claim than yours. If they want the gate, focus on performance risk and the hedge.

This is an example. Your plan uses your own week.

Make my plan

Also matched: sdet, test automation engineer, software development engineer in test. Research checked 2026-10-11. Page updated .