Essay reframes AI coding burnout as tool churn, not overload

An August 23 essay on paddo.dev argues that developer burnout on AI coding isn't a workload problem, it's a knowledge problem. The exhaustion comes from learning tools that expire before the lessons on them pay off.

The essay leans on real numbers to make its case. A Sonar survey of more than 1,100 developers found 96% don't fully trust code written by AI tools, and 48% say they always check it before shipping. Faros telemetry shows review queues collapsing under that checking load. Neither DORA nor Stack Overflow has published a 2026 survey, so the most-cited burnout research still describes someone using an assistant, not supervising an agent, per the essay.

Then it lists what expired in August alone. MCP's July 28 spec update removed sessions from the protocol. DeepSeek's flat pricing ended August 16. Claude Code shipped five releases in four days. The essay's fix: stop chasing every update, hold your toolchain steady, and let the models underneath do the changing.

That's a real cost most teams don't put in a budget line, relearning a tool is time not spent building. But the fix has a hole. MCP removing sessions from its spec isn't a model changing quietly underneath a stable tool, it's the protocol itself moving. Keep your tools still only works if the ground they sit on stays put, and the same essay's own examples say it isn't.

Watch whether the next MCP revision breaks something else. If protocol-level churn keeps outpacing model-level churn, the fix stops being freeze your tools and starts being nobody's tools are stable enough to freeze.