What works, what goes wrong, and how to tell the difference.
Tests are the task agents are best at, because the work checks itself. Flint writes the test, runs it, and reads the failure, so what comes back is verified rather than plausible. The trap is asking for coverage instead of behaviour: a hundred generated tests that assert nothing is worse than none.
The rule that matters more than the tool: give the agent something that can prove it succeeded. A task with a test, a build or a linter attached comes back verified. A task with none comes back plausible, and plausible is the expensive kind of wrong.
A desktop AI agent for Windows, built in Auckland, New Zealand. It reads your files, runs your terminal and drives your browser. Billed per token with a free monthly allowance and no subscription. See pricing, the FAQ, or who builds it.