Four columns expose the real difference.

Short video caption

A fair harness comparison uses trigger, pre-effect behavior, route coverage and outside readback for each runtime. Mark unknowns, then run harmless canaries. Documentation snapshot; no cross-product runtime test claimed. #EricFieldNotes

Full written post / accessible read

Teams often compare agent products by their instruction files and hook count. A useful comparison starts with the action you need to control: write, shell, browser, hosted tool or deployment.

For example, an IDE agent, cloud agent and inline completion can use different callbacks. A policy tested in one may leave another unmeasured. That is an evaluation gap, not proof the vendor is unsafe.

Put each product and runtime on a row. Column one: the hook trigger. Two: can it veto before the side effect? Three: which tool routes are actually covered? Four: what independent readback proves the action did or did not occur? Fill unknowns with UNKNOWN.

Run a harmless attempted write or deployment with a disposable identity on every route. Save the hook log and outside state. Select the system only after the controls you need are observed, because a neat configuration file cannot certify runtime coverage.

#EricFieldNotes

Evidence and boundary

On-screen boundary: PRODUCT COMPARISON METHOD. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.