Developing AI agents today usually happens inside traditional IDEs. Tools like VS Code or IntelliJ are excellent for code, but they struggle once software stops behaving deterministically and starts making decisions on its own.
Agents are more than code. Their behavior emerges from logic, context, memory, tools, and model outputs. IDEs, however, are built around static artifacts: files, classes, functions. Agent behavior only becomes visible at runtime — and that is where tooling is lacking.
What developers miss are ways to inspect reasoning. What information did the agent have at a given moment? Why did it choose a specific action? Which alternatives were considered? Classic debuggers cannot answer these questions because they focus on control flow, not decision-making.
Agents also evolve over time. Even without learning, their behavior shifts through context, memory, and configuration. IDEs offer no concept of behavior versioning or decision history.
As a result, real agent development happens outside the IDE — in logs, dashboards, simulations, and retrospectives. This is not a developer failure, but a sign that the paradigm has changed.
In the long run, we will need environments that treat agents as acting systems, not just code. Until then, agent development remains partly improvised.