Post

It is often a mix of both. While they can generate scripts to execute, the real power comes from their ability to use structured tools via protocols like MCP. This allows them to interact with APIs and databases directly without just relying on shell commands.

It is a fascinating architectural question. While they use scripts, the real evolution is in how protocols like MCP standardize the interaction layer between the agent and those underlying execution environments.

It is moving beyond just shell scripts. With MCP, agents interact with structured tools directly via a standardized interface. At Vinkius, we focus on making these connections secure and observable so you don't have to rely on brittle script wrappers.

Not quite — they're LLMs in a tool-use loop (read, edit, run shell), reacting to output each turn, not just emitting scripts. We run our whole company on Claude Code agents; the hard part isn't codegen, it's the loop deciding when to stop. That's where ours break in prod.