Gemini 3.5 Flash: evaluate an agent through its actions

Flash opens the family

In our retrospective May review, Google introduced Gemini 3.5 on May 19, 2026, releasing 3.5 Flash first, including through the Gemini API. The announcement separates this availability from Pro, which was still to come. Google presentation.

Look at a sequence of actions

For an agent working on a repository or finding information, a good final answer does not describe the whole job. It is useful to examine tool calls, selected data and evidence supporting the result. A known maintenance task offers a more revealing trial than an open-ended demonstration: the expected outcome can be defined and every proposed modification examined.

Define permission to act

Before comparing models, specify authorized actions to avoid confusing capability with access. Reading a file, editing a draft and publishing a change have different consequences. The relevant cost is the workflow that reaches an accepted result, including duration and retries. Vendor demonstrations suggest directions; they guarantee neither quality on a particular project nor operation without supervision.