Gartner listed six software engineering trends on November 18, for the first time including Harness engineering, Agent sandboxes and evaluation-driven development. Three of its predictions point to the same problem: AI can already write code, but enterprises have not yet built the processes needed to determine whether that code is fit for purpose.
What this is
This is not another round of competition over code-generation capabilities. It is the beginning of enterprises adding engineering systems to support AI’s role in development. Harness engineering adds a layer that defines how AI or software should run, be tested and be evaluated. An Agent sandbox is an isolated runtime that limits what an AI agent can access or operate. Evaluation-driven development requires teams to define output-quality criteria before executing a task.
All three address the same growing problem: it is becoming increasingly difficult to define when delivery is “complete.” Gartner predicts that by 2028, Harness engineering could reduce deployment defects by 60%, 70% of teams will adopt evaluation-driven development, and by 2030, 70% of enterprises will sandbox their agent development workflows. These figures are forecasts, not results that have already been achieved.
The other three trends involve AI-Native development platforms, on-premises AI and forward-deployed engineers. The first two change how applications are built and where they are deployed. The third places more engineers at customer sites, changing organizational and delivery structures.
Industry view
We believe the immediate priority is evaluation, sandboxing and Harness engineering—not replacing an entire development platform. Teams that already have testing and continuous integration capabilities can use fixed scripts, isolated environments and regression cases to turn “it looks fine to me” into repeatable acceptance testing.
The risks are equally clear. These methods add up-front configuration and maintenance costs. For teams with weak test coverage or prototypes that do not require long-term maintenance, rushing to build a complete system could slow delivery instead. Gartner’s percentages may also be affected by tool maturity, corporate compliance requirements and willingness to adopt. They should not be treated as procurement commitments or KPIs.
Impact on regular people
For enterprise IT: When selecting Agent tools, isolated execution, test logs and evaluation capabilities will matter more than raw generation speed. Teams with existing CI and automated testing will be better positioned to adopt this workflow.
For individual careers: When using AI to modify a real project, the minimum viable approach is to specify the directories it may modify, the tests that must pass and the checks required before submission, then continuously validate its output against 10 fixed test cases.
For the consumer market: As enterprises embed acceptance testing into AI development, users may encounter fewer deployment failures in the software they ultimately use. But if companies prioritize generation speed while neglecting testing and boundaries, the spread of AI coding could also amplify quality risks.