Both greater constraints and more introspective capabilities will let us more confidently move faster with coding agents.
Recently I’ve made a debugger for WebAssembly: you can step over each wasm instruction and see the memory and state at each point.
Because the debugger is a TUI, a coding agent can also use the debugger! I tell the agent to optimize using the debugger and it can benchmark before and after and study how the code it changes compiles and executes.
Here’s a rgb to hex wasm module:
echo 'rgb(101, 79, 240)' | npx @qip.dev/qipx qip.dev run text/rgb-to-hex.wasm