How the roll-up gate works
How the Agent Side-Effect & Blast-Radius gate works
This tool grades each proposed AI-agent action on reversibility and blast radius before it runs. It reads 6 actions and rolls them into one verdict — the worst, never the average — then a gate that can only make it worse.
The 6 actions
01A-04 · send_customer_emailfix firstno undo mechanism at all (IRREVERSIBLE) - narrow radius doesn't offset it - decided by reversibilityDO NOT AUTOMATE
02A-03 · update_crm_fieldreversible, but no enforced cap - blast radius UNBOUNDED - decided by blast radiusDO NOT AUTOMATE
03A-05 · apply_account_creditcompensating reversal only (COMPENSABLE), 5 subjects per callRUN WITH APPROVAL
04A-01 · issue_refunddeclared REVERSIBLE, but detect+execute takes 255 min against a 60-min undo window - demoted to COMPENSABLERUN WITH APPROVAL
05A-06 · tag_leadreversible and inside its window, but 200 subjects per call - blast radius WIDERUN WITH APPROVAL
06A-02 · issue_refund (monitored)reversible, undo finishes in 35 min inside the 60-min window, 1 subject per callRUN UNATTENDED
SAFE UNATTENDED
every action is reversible within its window and radius-capped
APPROVALS NEEDED
worst action is compensable-but-slow or radius-wide, not irreversible or unbounded
UNSAFE TO AUTOMATE
any action is irreversible, has no enforced cap, or can't finish inside its undo window
Undo-window gate · worsen-only
A true_undo action only counts as REVERSIBLE if detection plus execution time fits inside its declared undo window; if it doesn't, the action is demoted to COMPENSABLE regardless of the mechanism the agent declared.
UNSAFE TO AUTOMATEWhy: send_customer_email has no undo mechanism at all, so it's IRREVERSIBLE - and even though its blast radius is narrow (1 subject per call), reversibility alone is enough to make the whole portfolio unsafe to automate. 1 unattended - 3 with approval - 2 not automatable, of 6 actions - 17% could run unattended (context only).
Fix firstA-04, send_customer_email - add a real undo mechanism before this action can be automated; a narrow radius doesn't compensate for zero reversibility.
No single action is the crisis — the clustering is. The roll-up takes the worst, the gate escalates when trouble compounds, and it names the one thing to fix first. Grades the actions you describe. It is a working aid, not legal, security, or compliance advice, and it never scores, ranks, or assesses people.
RedHub AI