Testing in the Age of AI Agents
When code changes at the speed of thought, tests become less about verification and more about defining what should remain stable. A philosophy for testing when AI agents write and rewrite your code.
When code changes at the speed of thought, tests become less about verification and more about defining what should remain stable. A philosophy for testing when AI agents write and rewrite your code.
I ran benchmarks comparing CLI tools against MCP servers for browser automation. 13x more token efficient with CLI. Here's what I found.
Teaching an agent to use CLI tools isn't about writing perfect documentation. It's about creating a feedback loop where the tool teaches, the agent learns, and reflection builds institutional knowledge.
One year after MCP's launch, Anthropic published a solution to context overhead: filesystem-based tool discovery. The pattern looked familiar.