What do this year’s conference submissions reveal about where test automation is heading? As a member of the AutomationSTAR 2026 Programme Committee, Gomathi Ramalingam has been reviewing the ideas and experiences practitioners want to share. Here, she picks out six trends that stood out—and offers a practical starting point for each.
You can join her tutorial, AI Testing Dojo: Build, Break & Master Agents, at AutomationSTAR 2026.
Every year someone publishes a list of testing trends, and every year most of it is weather. It rolls over the industry, makes a lot of noise, and leaves your Tuesday unchanged. So when we read through the AutomationSTAR 2026 submissions on the programme committee, I was looking for something different: not what people are excited about, but what is already altering the day-to-day work of the teams behind the submissions.
Two numbers stuck with me. Around two-thirds of this year’s abstracts mention AI, LLMs or agents in some form. And ‘trust’ turned up in roughly one in five, which is a vocabulary shift I have never seen at this scale before.
Put those together and you have the story of 2026: the industry has stopped asking whether AI will write our tests and started asking how on earth we know when it has done it well.
Here are the six trends I think deserve your attention, each with one thing you can start this week.
1. Agentic testing is moving from demo to duty
For the last two years, ‘AI in testing’ mostly meant a chat window that produced plausible-looking test cases. That picture is changing. The pattern now showing up in real pipelines is a set of specialised agents with narrow jobs: a planner that explores the application and drafts a plan, a generator that turns the plan into executable code, and a healer that re-runs failures and repairs broken locators.
Playwright has formalised exactly this planner–generator–healer split and the Model Context Protocol (MCP) has become the standard plumbing that lets an agent see an accessibility tree rather than a wall of raw HTML.
The reason this matters is that decomposition is a hallucination-control strategy. An agent with one small, checkable job is far easier to audit than a monolith that “does testing”.
Do this: Pick one narrow, boring job in your pipeline, such as triaging failed runs or proposing locator fixes and give it to an agent with a human approval step. Do not start with “generate our regression suite”.
2. Trust has replaced coverage as the metric that matters
The World Quality Report 2025–26 found that 89% of organisations are piloting or deploying GenAI in quality engineering, yet only 15% have managed enterprise-wide adoption. The most telling figure is the barrier list: 60% cite hallucination and reliability as a blocker. Teams are not short of AI-generated tests. They are short of a reason to believe them.
A green tick from a test the model wrote, asserting a value the model also invented, is not evidence of anything. I have watched a beautifully formatted suite pass at 100% because every assertion was tautological. The work of a tester in 2026 is increasingly to be the accountability layer for output that was produced faster than any human could read it.
Do this: Add a quality gate before AI-generated tests are merged. Mine has six checks, but even three will change your outcomes: does each test assert something the code could plausibly get wrong, does it reference a real endpoint or element, and would it fail if the feature were deleted?
3. AI-generated code needs AI-aware testing
It is not only tests being generated. A large and growing share of production code now starts life in a model, and the trust gap is widening: in Stack Overflow’s 2025 developer survey, 84% of developers use or plan to use AI tools, yet 46% say they do not trust the accuracy of the output, up from 31% the year before. That code has characteristic failure modes. It handles the happy path elegantly, invents APIs that do not exist, and quietly drops edge cases that were never in the prompt.
This is good news for testers. Classic technique is back in fashion. Boundary values, negative paths, state transitions and idempotency checks are precisely the things a model skips when it is optimising for “looks correct”.
Do this: When a change is flagged as AI-assisted, run a short, explicit review checklist against it rather than the same one you use for human code. The checklist should include “which inputs did the prompt not mention?”
Be in the room for all the important conversations at AutomationSTAR 2026 – book your tickets now.
4. Flaky tests are finally being priced
For years flaky tests were tolerated as a tax. This year’s submissions and blog posts treat them as a cost line, and that reframing is overdue. Every retry, every “re-run and it passed”, every engineer who has learned to ignore red builds is a measurable drain on delivery. And in an agentic world it is worse: an agent that learns your suite is noisy will learn to retry, heal and paper over, which is exactly the behaviour you do not want automated.
Do this: Spend an hour producing a flaky test audit. List the tests that failed and then passed on retry in the last thirty days, with the minutes of pipeline time they consumed. Take the number to whoever owns the roadmap. It is a far more persuasive argument than “we need to fix tech debt”.
5. API-first testing is the natural habitat for AI
There is a reason so much of the serious agentic work is happening at the API layer rather than the UI. APIs are text. Specifications are text. Requests and responses are structured, deterministic and cheap to run thousands of times. A model reasoning over an OpenAPI document is operating in its comfort zone; a model reasoning over a rendered DOM is guessing at pixels.
Contract testing fits this beautifully. A contract is a machine-readable statement of intent, which means it can serve as both the prompt and the guardrail for generated tests.
Do this: If you are going to let a model generate tests anywhere, start with your APIs, feed it the spec rather than a description and measure hallucination rate per prompt. You will quickly discover that some ways of asking produce reliable tests and others produce fiction. Keep a scored library of the ones that work.
6. The tester’s skill stack is shifting, and it is not towards prompting
The tempting conclusion from all of the above is that testers need to become prompt engineers. I do not think that is quite right. The skills the 2026 submissions actually reward are older and harder: writing precise context, designing an experiment that can distinguish a good output from a convincing one, and thinking adversarially about a system that will confidently tell you it is fine.
That is testing. It always was. The difference is that the system under test now includes the tool doing the testing.
Do this: Give your team a low-stakes sandbox and a challenge: build a small testing agent, then break it. Nothing calibrates judgement about where AI helps and where it is dangerously wrong faster than watching your own agent fail. It is the format I am using for a hands-on tutorial at AutomationSTAR this November, and it is the approach I would recommend whether or not you attend.
The thread running through all six
Automation has always been about removing humans from the parts of the work that do not need them. What 2026 is clarifying is which parts those are. Generation is being automated. Judgement is not. The teams that will do well are the ones that treat every green tick from an AI as a claim to be verified, not a result to be celebrated.
Want to take these ideas further? Join us at AutomationSTAR in Antwerp on 4–5 November for practical talks, hands-on tutorials and conversations with the people putting them to work. Explore the programme, and find the sessions that tackle your team’s next challenge. Check out and book your AutomationSTAR tickets.
