Agentic AI needs more testers, not fewer – Unite.AI

Agentic AI is rewriting how software ships. Assistants write pull requests, agents open migrations, and review bots say a change is safe before a human runs out of coffee. The industry often frames this as a path to fewer people in terms of quality. That shot is backwards. When machines create more of the code base, the poor skill is not generation. It is an independent verification.

The 2025 Stack Overflow Developer Survey found that 84% of respondents use or plan to use AI tools during development, and 51% of professional developers use them daily. Positive sentiment has dropped to 60%, and more developers actively distrust the accuracy of AI (46%) than distrust it (33%). The top frustration, cited by 66%, was solutions that were almost right but not quite right. Google Cloud’s 2025 DORA research reported AI adoption at 90% among software development professionals, with more than 80% saying AI has improved productivity, but only 24% reporting a lot or a great deal of confidence. Utility without trust is precisely the condition that expands the tester’s work, not narrows it.

Speed ​​is not a test

I work as a software quality engineer focused on agent testing and CI ports for AI-created code. Internally, we have provided a PR reviewer and a test case generator. Both are drafting tools. Neither of them evaluates their homework. This rule is important because agents speak with certainty. Certainty is not correctness.

We learned this the hard way. In an AI-created PR review pilot, one edit cleared the CI in less than 4 minutes. Two days later, a downstream dashboard showed impossible temperatures. A scale reversal from Celsius to Fahrenheit occurred on a code path that the suite never used. The pipeline was green. The product was wrong. The green IC that has never asked itself the difficult question is theatre.

This incident fits into a larger pattern. CodeRabbit’s State of AI vs Human Code Generation report analyzed 470 open source GitHub pull requests (320 AI co-authored, 150 human-only). The AI ​​changes resulted in approximately 1.7 times more issues overall, with logic and correctness issues being 75% more common and up to 2.74 times higher security outcomes. Veracode’s 2026 GenAI Code Security Report tested patterns in security tasks and found an average security pass rate of 56%, little changed from 55% in the previous report, while approximately 44% of AI code generation tasks introduced a known vulnerability. The syntax is almost solved. Safe and correct behavior is not. Testers are the people who refuse to confuse these two.

Hallucinatory addictions are a testing problem

The generation also invents the offer. The USENIX Security 2025 study, We Have a Package for You!, evaluated package hallucinations in 16 LLM code generators and 576,000 samples. Commercial models hallucinated on average at least 5.2% of the time; open source models averaged 21.7%, with 205,474 unique hallucinated names. Persistent ghost packages create a ripe target for slop squat: Attackers register the made-up name and wait for a developer or agent to install it.

Dependency provenance checks belong to the same level as functional testing. If an agent invents a NuGet or npm package, the merge would have to fail before anyone could argue whether the unit testing was thorough. Provenance is not optional hygiene. It is a quality gate for change created by artificial intelligence.

Gates that adapt to agents

When we did an organization-wide upgrade from .NET 8 to .NET 10 in about two weeks with AI-assisted automation, I was the quality manager. The speed was real. So was the need for controls that did not believe the agent’s synthesis. The ports I run on the modified code include:

• Mutation score targets between 60% and 70% on touched logic, so the suite must kill actual mutants rather than just cover lines
• Property-based tests on high-stakes logic, where examples alone underspecify the behavior
• Dependency provenance checks with respect to slop squatting and hallucinated packages
• Duplication and abandonment signs signaling copy-paste thrash by officers
• Independently verify agent claims such as “migration completed,” because a claim is a guess until you get proof

The mutation score is not a vanity metric. Asks if your tests detect an intentional error. Coverage can increase while the suite does not yet reach the Celsius path. Property-based controls force the system model to stay on different ranges, not on a happy path. The provenance closes the package pantry. Abandonment and duplication capture the agent’s habit of rewriting the same idea in three ways. Together they form a posture: assume the agent is wrong until the suite proves otherwise. I keep an eye on the calibration line.

None of this replaces API testing, authentication testing, smoke testing, or UI automation. FlaUI and PowerShell still catch what the APIs are missing. Package creation via NuGet, MSI, and MSM still requires silent path installation and validation. Agent workflows amplify the volume. They don’t retire those disciplines.

More testers, different work

The fear is that agents will write the tests and humans will become optional. Officers can draft cases. They should not certify their own drafts. Stack Overflow respondents planning the future use of AI for testing still put human mistrust at the center: when asking for help from another person, the main reason was not to trust the AI’s answers. This instinct is the resource of the profession.

Here it is important to hire and enable people who have followed non-traditional paths. I transitioned into quality after a career in HR and data, completed a Master’s in Information Systems at UMBC in May 2024, and now spend my days making agent gates practical. Agentic AI will not wait for a single pedigree. It needs more people who can design proof, not fewer people who rubber stamp green construction.

The question about the product is simple. If agents ship faster than your audit might think, you won’t have any productivity benefits. You have a postponed accident. Invest in testers who treat every statement from an agent as suspicious, who link mutation and provenance to the IC, and who maintain independent judgment where the model is more fluid and less reliable. AI needs more testers, not fewer.