Agentic speed needs observable control.

Day 04 · 2026-10-01 · LinkedIn

Short video caption

AGENTS.md, CLAUDE.md and skills can describe policy; they do not themselves deny tool calls. A hook can intercept covered paths, while identity scope limits missed ones. Build a route-by-route proof map: action, runtime, failure behavior, observed denial and independent result. Codex, Claude Code and Cursor document different paths; verify yours with harmless canaries. #EricFieldNotes

Full written post / accessible read

Agents can multiply engineering output. But the gain disappears if every signoff rests on trust in a prompt. Instruction files shape behavior. The engineering question is which tool calls your policy intercepts, and on which runtime.

AGENTS files, CLAUDE files and skills can all tell a worker to stay inside policy. If the team tests only a shell hook, a hosted tool or directly typed command can fall outside that exact matcher. The cost is a safety claim the actual worker never earned.

Codex documents a local hook path that excludes hosted web tools. Claude Code has a separate event for directly typed slash-command expansion. Cursor's local user hooks do not load in cloud agents. Those are product-doc snapshots, not completed canary results.

Enumerate enabled routes, run a harmless deny canary on each deployed surface, and record what action occurred. Scope credentials separately and check the outcome independently. Do this because coverage is an observed property, not a feature checkbox.

#EricFieldNotes

Evidence and boundary

On-screen boundary: CROSS-RUNTIME CONTROL · DOC SNAPSHOT. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.