Coding agents need bounded tools and deliberate human checkpoints

The practical center is agent setup and tool authority. One harness story moves a Claude setup from a 1,800-line CLAUDE.md into scoped files, then into keystone-mcp, a Model Context Protocol server that exposes rules, skills, status, verification, and budget data to agents. Agent Joe applies control at the command layer: Rust-only work, no shell access, and fewer available actions. A prompt-template item shows a looser pattern, with parallel agents assigned separate goals and merged afterward, but it gives no evaluation beyond the claim that this can be faster and more detailed.

Two pieces focus on the operator, not the agent. One argues that instant AI responses remove the pauses that used to come from searching, waiting, and getting stuck. The reported failure mode is scope creep: a short task can turn into two hours of prompting, refactoring, and adding features before the original goal is checked. The proposed fix is plain: stop every 30 minutes and review the goal without the model. A second essay breaks engineering “taste” into product thinking, system thinking, and quality calibration. In that account, the human contribution is choosing the right problem, architecture, and level of rigor for the situation.

The blockchain item is conceptual, but it adds a useful adjacent theme: public systems depend on deciding which real-world claims deserve formal settlement. The essay treats Ethereum as an ancestor for programmable public state and names assets, identities, claims, receipts, and records as candidates for on-chain representation. It also argues that AI tools and the open web often create the language, evidence, and corrections before a protocol can harden a claim. No benchmarks or comparative performance results are reported.