Coding agents need safer senses and stricter cost accounting

peek-cli gives coding agents a narrow way to inspect web apps: screenshots from open browser tabs. The tool uses a Chrome extension and a local WebSocket daemon, then exposes simple CLI commands such as peeked start, peeked list, and peeked at <url>. Its safety claim is specific. The agent can request a screenshot, while browser control, script injection, clicks, and typing stay outside the interface. That makes it relevant for frontend debugging and localhost UI checks where visual state matters, but direct browser access would be too risky.

The AI forecast in the corpus is a cautionary cost model, not an experiment. It argues that current inference demand may depend on prices below true compute cost, with estimates ranging from $0.60–$0.70 paid per $1 of compute to below $0.10 in a pessimistic subsidy case. The software claim is also bounded: coding assistants may make a good engineer about 20–30% more productive on average, while 10x and 100x claims are treated as unsupported. The essay points to a common failure mode for code generation: large volumes of generated code can duplicate existing open-source functionality and add bugs, so output size is a weak proxy for productivity.